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_u32   ("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         loff_t len, vma_len;
 135         int ret;
 136         struct hstate *h = hstate_file(file);
 137
 138         /*
 139          * vma address alignment (but not the pgoff alignment) has
 140          * already been checked by prepare_hugepage_range.  If you add
 141          * any error returns here, do so after setting VM_HUGETLB, so
 142          * is_vm_hugetlb_page tests below unmap_region go the right
 143          * way when do_mmap unwinds (may be important on powerpc
 144          * and ia64).
 145          */
 146         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 147         vma->vm_ops = &hugetlb_vm_ops;
 148
 149         /*
 150          * page based offset in vm_pgoff could be sufficiently large to
 151          * overflow a loff_t when converted to byte offset.  This can
 152          * only happen on architectures where sizeof(loff_t) ==
 153          * sizeof(unsigned long).  So, only check in those instances.
 154          */
 155         if (sizeof(unsigned long) == sizeof(loff_t)) {
 156                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 157                         return -EINVAL;
 158         }
 159
 160         /* must be huge page aligned */
 161         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 162                 return -EINVAL;
 163
 164         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 165         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 166         /* check for overflow */
 167         if (len < vma_len)
 168                 return -EINVAL;
 169
 170         inode_lock(inode);
 171         file_accessed(file);
 172
 173         ret = -ENOMEM;
 174         if (!hugetlb_reserve_pages(inode,
 175                                 vma->vm_pgoff >> huge_page_order(h),
 176                                 len >> huge_page_shift(h), vma,
 177                                 vma->vm_flags))
 178                 goto out;
 179
 180         ret = 0;
 181         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 182                 i_size_write(inode, len);
 183 out:
 184         inode_unlock(inode);
 185
 186         return ret;
 187 }
 188
 189 /*
 190  * Called under mmap_write_lock(mm).
 191  */
 192
 193 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 194 static unsigned long
 195 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
 196                 unsigned long len, unsigned long pgoff, unsigned long flags)
 197 {
 198         struct hstate *h = hstate_file(file);
 199         struct vm_unmapped_area_info info;
 200
 201         info.flags = 0;
 202         info.length = len;
 203         info.low_limit = current->mm->mmap_base;
 204         info.high_limit = TASK_SIZE;
 205         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 206         info.align_offset = 0;
 207         return vm_unmapped_area(&info);
 208 }
 209
 210 static unsigned long
 211 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
 212                 unsigned long len, unsigned long pgoff, unsigned long flags)
 213 {
 214         struct hstate *h = hstate_file(file);
 215         struct vm_unmapped_area_info info;
 216
 217         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 218         info.length = len;
 219         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
 220         info.high_limit = current->mm->mmap_base;
 221         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 222         info.align_offset = 0;
 223         addr = vm_unmapped_area(&info);
 224
 225         /*
 226          * A failed mmap() very likely causes application failure,
 227          * so fall back to the bottom-up function here. This scenario
 228          * can happen with large stack limits and large mmap()
 229          * allocations.
 230          */
 231         if (unlikely(offset_in_page(addr))) {
 232                 VM_BUG_ON(addr != -ENOMEM);
 233                 info.flags = 0;
 234                 info.low_limit = current->mm->mmap_base;
 235                 info.high_limit = TASK_SIZE;
 236                 addr = vm_unmapped_area(&info);
 237         }
 238
 239         return addr;
 240 }
 241
 242 static unsigned long
 243 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 244                 unsigned long len, unsigned long pgoff, unsigned long flags)
 245 {
 246         struct mm_struct *mm = current->mm;
 247         struct vm_area_struct *vma;
 248         struct hstate *h = hstate_file(file);
 249
 250         if (len & ~huge_page_mask(h))
 251                 return -EINVAL;
 252         if (len > TASK_SIZE)
 253                 return -ENOMEM;
 254
 255         if (flags & MAP_FIXED) {
 256                 if (prepare_hugepage_range(file, addr, len))
 257                         return -EINVAL;
 258                 return addr;
 259         }
 260
 261         if (addr) {
 262                 addr = ALIGN(addr, huge_page_size(h));
 263                 vma = find_vma(mm, addr);
 264                 if (TASK_SIZE - len >= addr &&
 265                     (!vma || addr + len <= vm_start_gap(vma)))
 266                         return addr;
 267         }
 268
 269         /*
 270          * Use mm->get_unmapped_area value as a hint to use topdown routine.
 271          * If architectures have special needs, they should define their own
 272          * version of hugetlb_get_unmapped_area.
 273          */
 274         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
 275                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 276                                 pgoff, flags);
 277         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 278                         pgoff, flags);
 279 }
 280 #endif
 281
 282 static size_t
 283 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 284                         struct iov_iter *to, unsigned long size)
 285 {
 286         size_t copied = 0;
 287         int i, chunksize;
 288
 289         /* Find which 4k chunk and offset with in that chunk */
 290         i = offset >> PAGE_SHIFT;
 291         offset = offset & ~PAGE_MASK;
 292
 293         while (size) {
 294                 size_t n;
 295                 chunksize = PAGE_SIZE;
 296                 if (offset)
 297                         chunksize -= offset;
 298                 if (chunksize > size)
 299                         chunksize = size;
 300                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 301                 copied += n;
 302                 if (n != chunksize)
 303                         return copied;
 304                 offset = 0;
 305                 size -= chunksize;
 306                 i++;
 307         }
 308         return copied;
 309 }
 310
 311 /*
 312  * Support for read() - Find the page attached to f_mapping and copy out the
 313  * data. Its *very* similar to generic_file_buffered_read(), we can't use that
 314  * since it has PAGE_SIZE assumptions.
 315  */
 316 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 317 {
 318         struct file *file = iocb->ki_filp;
 319         struct hstate *h = hstate_file(file);
 320         struct address_space *mapping = file->f_mapping;
 321         struct inode *inode = mapping->host;
 322         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 323         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 324         unsigned long end_index;
 325         loff_t isize;
 326         ssize_t retval = 0;
 327
 328         while (iov_iter_count(to)) {
 329                 struct page *page;
 330                 size_t nr, copied;
 331
 332                 /* nr is the maximum number of bytes to copy from this page */
 333                 nr = huge_page_size(h);
 334                 isize = i_size_read(inode);
 335                 if (!isize)
 336                         break;
 337                 end_index = (isize - 1) >> huge_page_shift(h);
 338                 if (index > end_index)
 339                         break;
 340                 if (index == end_index) {
 341                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 342                         if (nr <= offset)
 343                                 break;
 344                 }
 345                 nr = nr - offset;
 346
 347                 /* Find the page */
 348                 page = find_lock_page(mapping, index);
 349                 if (unlikely(page == NULL)) {
 350                         /*
 351                          * We have a HOLE, zero out the user-buffer for the
 352                          * length of the hole or request.
 353                          */
 354                         copied = iov_iter_zero(nr, to);
 355                 } else {
 356                         unlock_page(page);
 357
 358                         /*
 359                          * We have the page, copy it to user space buffer.
 360                          */
 361                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 362                         put_page(page);
 363                 }
 364                 offset += copied;
 365                 retval += copied;
 366                 if (copied != nr && iov_iter_count(to)) {
 367                         if (!retval)
 368                                 retval = -EFAULT;
 369                         break;
 370                 }
 371                 index += offset >> huge_page_shift(h);
 372                 offset &= ~huge_page_mask(h);
 373         }
 374         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 375         return retval;
 376 }
 377
 378 static int hugetlbfs_write_begin(struct file *file,
 379                         struct address_space *mapping,
 380                         loff_t pos, unsigned len, unsigned flags,
 381                         struct page **pagep, void **fsdata)
 382 {
 383         return -EINVAL;
 384 }
 385
 386 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 387                         loff_t pos, unsigned len, unsigned copied,
 388                         struct page *page, void *fsdata)
 389 {
 390         BUG();
 391         return -EINVAL;
 392 }
 393
 394 static void remove_huge_page(struct page *page)
 395 {
 396         ClearPageDirty(page);
 397         ClearPageUptodate(page);
 398         delete_from_page_cache(page);
 399 }
 400
 401 static void
 402 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
 403 {
 404         struct vm_area_struct *vma;
 405
 406         /*
 407          * end == 0 indicates that the entire range after
 408          * start should be unmapped.
 409          */
 410         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 411                 unsigned long v_offset;
 412                 unsigned long v_end;
 413
 414                 /*
 415                  * Can the expression below overflow on 32-bit arches?
 416                  * No, because the interval tree returns us only those vmas
 417                  * which overlap the truncated area starting at pgoff,
 418                  * and no vma on a 32-bit arch can span beyond the 4GB.
 419                  */
 420                 if (vma->vm_pgoff < start)
 421                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 422                 else
 423                         v_offset = 0;
 424
 425                 if (!end)
 426                         v_end = vma->vm_end;
 427                 else {
 428                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 429                                                         + vma->vm_start;
 430                         if (v_end > vma->vm_end)
 431                                 v_end = vma->vm_end;
 432                 }
 433
 434                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 435                                                                         NULL);
 436         }
 437 }
 438
 439 /*
 440  * remove_inode_hugepages handles two distinct cases: truncation and hole
 441  * punch.  There are subtle differences in operation for each case.
 442  *
 443  * truncation is indicated by end of range being LLONG_MAX
 444  *      In this case, we first scan the range and release found pages.
 445  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
 446  *      maps and global counts.  Page faults can not race with truncation
 447  *      in this routine.  hugetlb_no_page() holds i_mmap_rwsem and prevents
 448  *      page faults in the truncated range by checking i_size.  i_size is
 449  *      modified while holding i_mmap_rwsem.
 450  * hole punch is indicated if end is not LLONG_MAX
 451  *      In the hole punch case we scan the range and release found pages.
 452  *      Only when releasing a page is the associated region/reserve map
 453  *      deleted.  The region/reserve map for ranges without associated
 454  *      pages are not modified.  Page faults can race with hole punch.
 455  *      This is indicated if we find a mapped page.
 456  * Note: If the passed end of range value is beyond the end of file, but
 457  * not LLONG_MAX this routine still performs a hole punch operation.
 458  */
 459 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 460                                    loff_t lend)
 461 {
 462         struct hstate *h = hstate_inode(inode);
 463         struct address_space *mapping = &inode->i_data;
 464         const pgoff_t start = lstart >> huge_page_shift(h);
 465         const pgoff_t end = lend >> huge_page_shift(h);
 466         struct vm_area_struct pseudo_vma;
 467         struct pagevec pvec;
 468         pgoff_t next, index;
 469         int i, freed = 0;
 470         bool truncate_op = (lend == LLONG_MAX);
 471
 472         vma_init(&pseudo_vma, current->mm);
 473         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 474         pagevec_init(&pvec);
 475         next = start;
 476         while (next < end) {
 477                 /*
 478                  * When no more pages are found, we are done.
 479                  */
 480                 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
 481                         break;
 482
 483                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 484                         struct page *page = pvec.pages[i];
 485                         u32 hash;
 486
 487                         index = page->index;
 488                         hash = hugetlb_fault_mutex_hash(mapping, index);
 489                         if (!truncate_op) {
 490                                 /*
 491                                  * Only need to hold the fault mutex in the
 492                                  * hole punch case.  This prevents races with
 493                                  * page faults.  Races are not possible in the
 494                                  * case of truncation.
 495                                  */
 496                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 497                         }
 498
 499                         /*
 500                          * If page is mapped, it was faulted in after being
 501                          * unmapped in caller.  Unmap (again) now after taking
 502                          * the fault mutex.  The mutex will prevent faults
 503                          * until we finish removing the page.
 504                          *
 505                          * This race can only happen in the hole punch case.
 506                          * Getting here in a truncate operation is a bug.
 507                          */
 508                         if (unlikely(page_mapped(page))) {
 509                                 BUG_ON(truncate_op);
 510
 511                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 512                                 i_mmap_lock_write(mapping);
 513                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 514                                 hugetlb_vmdelete_list(&mapping->i_mmap,
 515                                         index * pages_per_huge_page(h),
 516                                         (index + 1) * pages_per_huge_page(h));
 517                                 i_mmap_unlock_write(mapping);
 518                         }
 519
 520                         lock_page(page);
 521                         /*
 522                          * We must free the huge page and remove from page
 523                          * cache (remove_huge_page) BEFORE removing the
 524                          * region/reserve map (hugetlb_unreserve_pages).  In
 525                          * rare out of memory conditions, removal of the
 526                          * region/reserve map could fail. Correspondingly,
 527                          * the subpool and global reserve usage count can need
 528                          * to be adjusted.
 529                          */
 530                         VM_BUG_ON(PagePrivate(page));
 531                         remove_huge_page(page);
 532                         freed++;
 533                         if (!truncate_op) {
 534                                 if (unlikely(hugetlb_unreserve_pages(inode,
 535                                                         index, index + 1, 1)))
 536                                         hugetlb_fix_reserve_counts(inode);
 537                         }
 538
 539                         unlock_page(page);
 540                         if (!truncate_op)
 541                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 542                 }
 543                 huge_pagevec_release(&pvec);
 544                 cond_resched();
 545         }
 546
 547         if (truncate_op)
 548                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 549 }
 550
 551 static void hugetlbfs_evict_inode(struct inode *inode)
 552 {
 553         struct resv_map *resv_map;
 554
 555         remove_inode_hugepages(inode, 0, LLONG_MAX);
 556
 557         /*
 558          * Get the resv_map from the address space embedded in the inode.
 559          * This is the address space which points to any resv_map allocated
 560          * at inode creation time.  If this is a device special inode,
 561          * i_mapping may not point to the original address space.
 562          */
 563         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
 564         /* Only regular and link inodes have associated reserve maps */
 565         if (resv_map)
 566                 resv_map_release(&resv_map->refs);
 567         clear_inode(inode);
 568 }
 569
 570 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 571 {
 572         pgoff_t pgoff;
 573         struct address_space *mapping = inode->i_mapping;
 574         struct hstate *h = hstate_inode(inode);
 575
 576         BUG_ON(offset & ~huge_page_mask(h));
 577         pgoff = offset >> PAGE_SHIFT;
 578
 579         i_mmap_lock_write(mapping);
 580         i_size_write(inode, offset);
 581         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 582                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 583         i_mmap_unlock_write(mapping);
 584         remove_inode_hugepages(inode, offset, LLONG_MAX);
 585 }
 586
 587 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 588 {
 589         struct hstate *h = hstate_inode(inode);
 590         loff_t hpage_size = huge_page_size(h);
 591         loff_t hole_start, hole_end;
 592
 593         /*
 594          * For hole punch round up the beginning offset of the hole and
 595          * round down the end.
 596          */
 597         hole_start = round_up(offset, hpage_size);
 598         hole_end = round_down(offset + len, hpage_size);
 599
 600         if (hole_end > hole_start) {
 601                 struct address_space *mapping = inode->i_mapping;
 602                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 603
 604                 inode_lock(inode);
 605
 606                 /* protected by i_rwsem */
 607                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 608                         inode_unlock(inode);
 609                         return -EPERM;
 610                 }
 611
 612                 i_mmap_lock_write(mapping);
 613                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 614                         hugetlb_vmdelete_list(&mapping->i_mmap,
 615                                                 hole_start >> PAGE_SHIFT,
 616                                                 hole_end  >> PAGE_SHIFT);
 617                 i_mmap_unlock_write(mapping);
 618                 remove_inode_hugepages(inode, hole_start, hole_end);
 619                 inode_unlock(inode);
 620         }
 621
 622         return 0;
 623 }
 624
 625 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 626                                 loff_t len)
 627 {
 628         struct inode *inode = file_inode(file);
 629         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 630         struct address_space *mapping = inode->i_mapping;
 631         struct hstate *h = hstate_inode(inode);
 632         struct vm_area_struct pseudo_vma;
 633         struct mm_struct *mm = current->mm;
 634         loff_t hpage_size = huge_page_size(h);
 635         unsigned long hpage_shift = huge_page_shift(h);
 636         pgoff_t start, index, end;
 637         int error;
 638         u32 hash;
 639
 640         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 641                 return -EOPNOTSUPP;
 642
 643         if (mode & FALLOC_FL_PUNCH_HOLE)
 644                 return hugetlbfs_punch_hole(inode, offset, len);
 645
 646         /*
 647          * Default preallocate case.
 648          * For this range, start is rounded down and end is rounded up
 649          * as well as being converted to page offsets.
 650          */
 651         start = offset >> hpage_shift;
 652         end = (offset + len + hpage_size - 1) >> hpage_shift;
 653
 654         inode_lock(inode);
 655
 656         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 657         error = inode_newsize_ok(inode, offset + len);
 658         if (error)
 659                 goto out;
 660
 661         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 662                 error = -EPERM;
 663                 goto out;
 664         }
 665
 666         /*
 667          * Initialize a pseudo vma as this is required by the huge page
 668          * allocation routines.  If NUMA is configured, use page index
 669          * as input to create an allocation policy.
 670          */
 671         vma_init(&pseudo_vma, mm);
 672         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 673         pseudo_vma.vm_file = file;
 674
 675         for (index = start; index < end; index++) {
 676                 /*
 677                  * This is supposed to be the vaddr where the page is being
 678                  * faulted in, but we have no vaddr here.
 679                  */
 680                 struct page *page;
 681                 unsigned long addr;
 682
 683                 cond_resched();
 684
 685                 /*
 686                  * fallocate(2) manpage permits EINTR; we may have been
 687                  * interrupted because we are using up too much memory.
 688                  */
 689                 if (signal_pending(current)) {
 690                         error = -EINTR;
 691                         break;
 692                 }
 693
 694                 /* Set numa allocation policy based on index */
 695                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 696
 697                 /* addr is the offset within the file (zero based) */
 698                 addr = index * hpage_size;
 699
 700                 /*
 701                  * fault mutex taken here, protects against fault path
 702                  * and hole punch.  inode_lock previously taken protects
 703                  * against truncation.
 704                  */
 705                 hash = hugetlb_fault_mutex_hash(mapping, index);
 706                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 707
 708                 /* See if already present in mapping to avoid alloc/free */
 709                 page = find_get_page(mapping, index);
 710                 if (page) {
 711                         put_page(page);
 712                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 713                         hugetlb_drop_vma_policy(&pseudo_vma);
 714                         continue;
 715                 }
 716
 717                 /*
 718                  * Allocate page without setting the avoid_reserve argument.
 719                  * There certainly are no reserves associated with the
 720                  * pseudo_vma.  However, there could be shared mappings with
 721                  * reserves for the file at the inode level.  If we fallocate
 722                  * pages in these areas, we need to consume the reserves
 723                  * to keep reservation accounting consistent.
 724                  */
 725                 page = alloc_huge_page(&pseudo_vma, addr, 0);
 726                 hugetlb_drop_vma_policy(&pseudo_vma);
 727                 if (IS_ERR(page)) {
 728                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 729                         error = PTR_ERR(page);
 730                         goto out;
 731                 }
 732                 clear_huge_page(page, addr, pages_per_huge_page(h));
 733                 __SetPageUptodate(page);
 734                 error = huge_add_to_page_cache(page, mapping, index);
 735                 if (unlikely(error)) {
 736                         put_page(page);
 737                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 738                         goto out;
 739                 }
 740
 741                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 742
 743                 SetHPageMigratable(page);
 744                 /*
 745                  * unlock_page because locked by add_to_page_cache()
 746                  * put_page() due to reference from alloc_huge_page()
 747                  */
 748                 unlock_page(page);
 749                 put_page(page);
 750         }
 751
 752         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 753                 i_size_write(inode, offset + len);
 754         inode->i_ctime = current_time(inode);
 755 out:
 756         inode_unlock(inode);
 757         return error;
 758 }
 759
 760 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
 761                              struct dentry *dentry, struct iattr *attr)
 762 {
 763         struct inode *inode = d_inode(dentry);
 764         struct hstate *h = hstate_inode(inode);
 765         int error;
 766         unsigned int ia_valid = attr->ia_valid;
 767         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 768
 769         error = setattr_prepare(&init_user_ns, dentry, attr);
 770         if (error)
 771                 return error;
 772
 773         if (ia_valid & ATTR_SIZE) {
 774                 loff_t oldsize = inode->i_size;
 775                 loff_t newsize = attr->ia_size;
 776
 777                 if (newsize & ~huge_page_mask(h))
 778                         return -EINVAL;
 779                 /* protected by i_rwsem */
 780                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 781                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 782                         return -EPERM;
 783                 hugetlb_vmtruncate(inode, newsize);
 784         }
 785
 786         setattr_copy(&init_user_ns, inode, attr);
 787         mark_inode_dirty(inode);
 788         return 0;
 789 }
 790
 791 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 792                                         struct hugetlbfs_fs_context *ctx)
 793 {
 794         struct inode *inode;
 795
 796         inode = new_inode(sb);
 797         if (inode) {
 798                 inode->i_ino = get_next_ino();
 799                 inode->i_mode = S_IFDIR | ctx->mode;
 800                 inode->i_uid = ctx->uid;
 801                 inode->i_gid = ctx->gid;
 802                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 803                 inode->i_op = &hugetlbfs_dir_inode_operations;
 804                 inode->i_fop = &simple_dir_operations;
 805                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 806                 inc_nlink(inode);
 807                 lockdep_annotate_inode_mutex_key(inode);
 808         }
 809         return inode;
 810 }
 811
 812 /*
 813  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 814  * be taken from reclaim -- unlike regular filesystems. This needs an
 815  * annotation because huge_pmd_share() does an allocation under hugetlb's
 816  * i_mmap_rwsem.
 817  */
 818 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 819
 820 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 821                                         struct inode *dir,
 822                                         umode_t mode, dev_t dev)
 823 {
 824         struct inode *inode;
 825         struct resv_map *resv_map = NULL;
 826
 827         /*
 828          * Reserve maps are only needed for inodes that can have associated
 829          * page allocations.
 830          */
 831         if (S_ISREG(mode) || S_ISLNK(mode)) {
 832                 resv_map = resv_map_alloc();
 833                 if (!resv_map)
 834                         return NULL;
 835         }
 836
 837         inode = new_inode(sb);
 838         if (inode) {
 839                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 840
 841                 inode->i_ino = get_next_ino();
 842                 inode_init_owner(&init_user_ns, inode, dir, mode);
 843                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 844                                 &hugetlbfs_i_mmap_rwsem_key);
 845                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 846                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 847                 inode->i_mapping->private_data = resv_map;
 848                 info->seals = F_SEAL_SEAL;
 849                 switch (mode & S_IFMT) {
 850                 default:
 851                         init_special_inode(inode, mode, dev);
 852                         break;
 853                 case S_IFREG:
 854                         inode->i_op = &hugetlbfs_inode_operations;
 855                         inode->i_fop = &hugetlbfs_file_operations;
 856                         break;
 857                 case S_IFDIR:
 858                         inode->i_op = &hugetlbfs_dir_inode_operations;
 859                         inode->i_fop = &simple_dir_operations;
 860
 861                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 862                         inc_nlink(inode);
 863                         break;
 864                 case S_IFLNK:
 865                         inode->i_op = &page_symlink_inode_operations;
 866                         inode_nohighmem(inode);
 867                         break;
 868                 }
 869                 lockdep_annotate_inode_mutex_key(inode);
 870         } else {
 871                 if (resv_map)
 872                         kref_put(&resv_map->refs, resv_map_release);
 873         }
 874
 875         return inode;
 876 }
 877
 878 /*
 879  * File creation. Allocate an inode, and we're done..
 880  */
 881 static int do_hugetlbfs_mknod(struct inode *dir,
 882                         struct dentry *dentry,
 883                         umode_t mode,
 884                         dev_t dev,
 885                         bool tmpfile)
 886 {
 887         struct inode *inode;
 888         int error = -ENOSPC;
 889
 890         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 891         if (inode) {
 892                 dir->i_ctime = dir->i_mtime = current_time(dir);
 893                 if (tmpfile) {
 894                         d_tmpfile(dentry, inode);
 895                 } else {
 896                         d_instantiate(dentry, inode);
 897                         dget(dentry);/* Extra count - pin the dentry in core */
 898                 }
 899                 error = 0;
 900         }
 901         return error;
 902 }
 903
 904 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
 905                            struct dentry *dentry, umode_t mode, dev_t dev)
 906 {
 907         return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
 908 }
 909
 910 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
 911                            struct dentry *dentry, umode_t mode)
 912 {
 913         int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
 914                                      mode | S_IFDIR, 0);
 915         if (!retval)
 916                 inc_nlink(dir);
 917         return retval;
 918 }
 919
 920 static int hugetlbfs_create(struct user_namespace *mnt_userns,
 921                             struct inode *dir, struct dentry *dentry,
 922                             umode_t mode, bool excl)
 923 {
 924         return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
 925 }
 926
 927 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
 928                              struct inode *dir, struct dentry *dentry,
 929                              umode_t mode)
 930 {
 931         return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
 932 }
 933
 934 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
 935                              struct inode *dir, struct dentry *dentry,
 936                              const char *symname)
 937 {
 938         struct inode *inode;
 939         int error = -ENOSPC;
 940
 941         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 942         if (inode) {
 943                 int l = strlen(symname)+1;
 944                 error = page_symlink(inode, symname, l);
 945                 if (!error) {
 946                         d_instantiate(dentry, inode);
 947                         dget(dentry);
 948                 } else
 949                         iput(inode);
 950         }
 951         dir->i_ctime = dir->i_mtime = current_time(dir);
 952
 953         return error;
 954 }
 955
 956 static int hugetlbfs_migrate_page(struct address_space *mapping,
 957                                 struct page *newpage, struct page *page,
 958                                 enum migrate_mode mode)
 959 {
 960         int rc;
 961
 962         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 963         if (rc != MIGRATEPAGE_SUCCESS)
 964                 return rc;
 965
 966         if (hugetlb_page_subpool(page)) {
 967                 hugetlb_set_page_subpool(newpage, hugetlb_page_subpool(page));
 968                 hugetlb_set_page_subpool(page, NULL);
 969         }
 970
 971         if (mode != MIGRATE_SYNC_NO_COPY)
 972                 migrate_page_copy(newpage, page);
 973         else
 974                 migrate_page_states(newpage, page);
 975
 976         return MIGRATEPAGE_SUCCESS;
 977 }
 978
 979 static int hugetlbfs_error_remove_page(struct address_space *mapping,
 980                                 struct page *page)
 981 {
 982         struct inode *inode = mapping->host;
 983         pgoff_t index = page->index;
 984
 985         remove_huge_page(page);
 986         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
 987                 hugetlb_fix_reserve_counts(inode);
 988
 989         return 0;
 990 }
 991
 992 /*
 993  * Display the mount options in /proc/mounts.
 994  */
 995 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
 996 {
 997         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
 998         struct hugepage_subpool *spool = sbinfo->spool;
 999         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1000         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1001         char mod;
1002
1003         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1004                 seq_printf(m, ",uid=%u",
1005                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1006         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1007                 seq_printf(m, ",gid=%u",
1008                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1009         if (sbinfo->mode != 0755)
1010                 seq_printf(m, ",mode=%o", sbinfo->mode);
1011         if (sbinfo->max_inodes != -1)
1012                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1013
1014         hpage_size /= 1024;
1015         mod = 'K';
1016         if (hpage_size >= 1024) {
1017                 hpage_size /= 1024;
1018                 mod = 'M';
1019         }
1020         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1021         if (spool) {
1022                 if (spool->max_hpages != -1)
1023                         seq_printf(m, ",size=%llu",
1024                                    (unsigned long long)spool->max_hpages << hpage_shift);
1025                 if (spool->min_hpages != -1)
1026                         seq_printf(m, ",min_size=%llu",
1027                                    (unsigned long long)spool->min_hpages << hpage_shift);
1028         }
1029         return 0;
1030 }
1031
1032 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1033 {
1034         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1035         struct hstate *h = hstate_inode(d_inode(dentry));
1036
1037         buf->f_type = HUGETLBFS_MAGIC;
1038         buf->f_bsize = huge_page_size(h);
1039         if (sbinfo) {
1040                 spin_lock(&sbinfo->stat_lock);
1041                 /* If no limits set, just report 0 for max/free/used
1042                  * blocks, like simple_statfs() */
1043                 if (sbinfo->spool) {
1044                         long free_pages;
1045
1046                         spin_lock(&sbinfo->spool->lock);
1047                         buf->f_blocks = sbinfo->spool->max_hpages;
1048                         free_pages = sbinfo->spool->max_hpages
1049                                 - sbinfo->spool->used_hpages;
1050                         buf->f_bavail = buf->f_bfree = free_pages;
1051                         spin_unlock(&sbinfo->spool->lock);
1052                         buf->f_files = sbinfo->max_inodes;
1053                         buf->f_ffree = sbinfo->free_inodes;
1054                 }
1055                 spin_unlock(&sbinfo->stat_lock);
1056         }
1057         buf->f_namelen = NAME_MAX;
1058         return 0;
1059 }
1060
1061 static void hugetlbfs_put_super(struct super_block *sb)
1062 {
1063         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1064
1065         if (sbi) {
1066                 sb->s_fs_info = NULL;
1067
1068                 if (sbi->spool)
1069                         hugepage_put_subpool(sbi->spool);
1070
1071                 kfree(sbi);
1072         }
1073 }
1074
1075 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1076 {
1077         if (sbinfo->free_inodes >= 0) {
1078                 spin_lock(&sbinfo->stat_lock);
1079                 if (unlikely(!sbinfo->free_inodes)) {
1080                         spin_unlock(&sbinfo->stat_lock);
1081                         return 0;
1082                 }
1083                 sbinfo->free_inodes--;
1084                 spin_unlock(&sbinfo->stat_lock);
1085         }
1086
1087         return 1;
1088 }
1089
1090 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1091 {
1092         if (sbinfo->free_inodes >= 0) {
1093                 spin_lock(&sbinfo->stat_lock);
1094                 sbinfo->free_inodes++;
1095                 spin_unlock(&sbinfo->stat_lock);
1096         }
1097 }
1098
1099
1100 static struct kmem_cache *hugetlbfs_inode_cachep;
1101
1102 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1103 {
1104         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1105         struct hugetlbfs_inode_info *p;
1106
1107         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1108                 return NULL;
1109         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1110         if (unlikely(!p)) {
1111                 hugetlbfs_inc_free_inodes(sbinfo);
1112                 return NULL;
1113         }
1114
1115         /*
1116          * Any time after allocation, hugetlbfs_destroy_inode can be called
1117          * for the inode.  mpol_free_shared_policy is unconditionally called
1118          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1119          * in case of a quick call to destroy.
1120          *
1121          * Note that the policy is initialized even if we are creating a
1122          * private inode.  This simplifies hugetlbfs_destroy_inode.
1123          */
1124         mpol_shared_policy_init(&p->policy, NULL);
1125
1126         return &p->vfs_inode;
1127 }
1128
1129 static void hugetlbfs_free_inode(struct inode *inode)
1130 {
1131         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1132 }
1133
1134 static void hugetlbfs_destroy_inode(struct inode *inode)
1135 {
1136         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1137         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1138 }
1139
1140 static const struct address_space_operations hugetlbfs_aops = {
1141         .write_begin    = hugetlbfs_write_begin,
1142         .write_end      = hugetlbfs_write_end,
1143         .set_page_dirty =  __set_page_dirty_no_writeback,
1144         .migratepage    = hugetlbfs_migrate_page,
1145         .error_remove_page      = hugetlbfs_error_remove_page,
1146 };
1147
1148
1149 static void init_once(void *foo)
1150 {
1151         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1152
1153         inode_init_once(&ei->vfs_inode);
1154 }
1155
1156 const struct file_operations hugetlbfs_file_operations = {
1157         .read_iter              = hugetlbfs_read_iter,
1158         .mmap                   = hugetlbfs_file_mmap,
1159         .fsync                  = noop_fsync,
1160         .get_unmapped_area      = hugetlb_get_unmapped_area,
1161         .llseek                 = default_llseek,
1162         .fallocate              = hugetlbfs_fallocate,
1163 };
1164
1165 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1166         .create         = hugetlbfs_create,
1167         .lookup         = simple_lookup,
1168         .link           = simple_link,
1169         .unlink         = simple_unlink,
1170         .symlink        = hugetlbfs_symlink,
1171         .mkdir          = hugetlbfs_mkdir,
1172         .rmdir          = simple_rmdir,
1173         .mknod          = hugetlbfs_mknod,
1174         .rename         = simple_rename,
1175         .setattr        = hugetlbfs_setattr,
1176         .tmpfile        = hugetlbfs_tmpfile,
1177 };
1178
1179 static const struct inode_operations hugetlbfs_inode_operations = {
1180         .setattr        = hugetlbfs_setattr,
1181 };
1182
1183 static const struct super_operations hugetlbfs_ops = {
1184         .alloc_inode    = hugetlbfs_alloc_inode,
1185         .free_inode     = hugetlbfs_free_inode,
1186         .destroy_inode  = hugetlbfs_destroy_inode,
1187         .evict_inode    = hugetlbfs_evict_inode,
1188         .statfs         = hugetlbfs_statfs,
1189         .put_super      = hugetlbfs_put_super,
1190         .show_options   = hugetlbfs_show_options,
1191 };
1192
1193 /*
1194  * Convert size option passed from command line to number of huge pages
1195  * in the pool specified by hstate.  Size option could be in bytes
1196  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1197  */
1198 static long
1199 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1200                          enum hugetlbfs_size_type val_type)
1201 {
1202         if (val_type == NO_SIZE)
1203                 return -1;
1204
1205         if (val_type == SIZE_PERCENT) {
1206                 size_opt <<= huge_page_shift(h);
1207                 size_opt *= h->max_huge_pages;
1208                 do_div(size_opt, 100);
1209         }
1210
1211         size_opt >>= huge_page_shift(h);
1212         return size_opt;
1213 }
1214
1215 /*
1216  * Parse one mount parameter.
1217  */
1218 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1219 {
1220         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1221         struct fs_parse_result result;
1222         char *rest;
1223         unsigned long ps;
1224         int opt;
1225
1226         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1227         if (opt < 0)
1228                 return opt;
1229
1230         switch (opt) {
1231         case Opt_uid:
1232                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1233                 if (!uid_valid(ctx->uid))
1234                         goto bad_val;
1235                 return 0;
1236
1237         case Opt_gid:
1238                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1239                 if (!gid_valid(ctx->gid))
1240                         goto bad_val;
1241                 return 0;
1242
1243         case Opt_mode:
1244                 ctx->mode = result.uint_32 & 01777U;
1245                 return 0;
1246
1247         case Opt_size:
1248                 /* memparse() will accept a K/M/G without a digit */
1249                 if (!isdigit(param->string[0]))
1250                         goto bad_val;
1251                 ctx->max_size_opt = memparse(param->string, &rest);
1252                 ctx->max_val_type = SIZE_STD;
1253                 if (*rest == '%')
1254                         ctx->max_val_type = SIZE_PERCENT;
1255                 return 0;
1256
1257         case Opt_nr_inodes:
1258                 /* memparse() will accept a K/M/G without a digit */
1259                 if (!isdigit(param->string[0]))
1260                         goto bad_val;
1261                 ctx->nr_inodes = memparse(param->string, &rest);
1262                 return 0;
1263
1264         case Opt_pagesize:
1265                 ps = memparse(param->string, &rest);
1266                 ctx->hstate = size_to_hstate(ps);
1267                 if (!ctx->hstate) {
1268                         pr_err("Unsupported page size %lu MB\n", ps >> 20);
1269                         return -EINVAL;
1270                 }
1271                 return 0;
1272
1273         case Opt_min_size:
1274                 /* memparse() will accept a K/M/G without a digit */
1275                 if (!isdigit(param->string[0]))
1276                         goto bad_val;
1277                 ctx->min_size_opt = memparse(param->string, &rest);
1278                 ctx->min_val_type = SIZE_STD;
1279                 if (*rest == '%')
1280                         ctx->min_val_type = SIZE_PERCENT;
1281                 return 0;
1282
1283         default:
1284                 return -EINVAL;
1285         }
1286
1287 bad_val:
1288         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1289                       param->string, param->key);
1290 }
1291
1292 /*
1293  * Validate the parsed options.
1294  */
1295 static int hugetlbfs_validate(struct fs_context *fc)
1296 {
1297         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1298
1299         /*
1300          * Use huge page pool size (in hstate) to convert the size
1301          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1302          */
1303         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1304                                                    ctx->max_size_opt,
1305                                                    ctx->max_val_type);
1306         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1307                                                    ctx->min_size_opt,
1308                                                    ctx->min_val_type);
1309
1310         /*
1311          * If max_size was specified, then min_size must be smaller
1312          */
1313         if (ctx->max_val_type > NO_SIZE &&
1314             ctx->min_hpages > ctx->max_hpages) {
1315                 pr_err("Minimum size can not be greater than maximum size\n");
1316                 return -EINVAL;
1317         }
1318
1319         return 0;
1320 }
1321
1322 static int
1323 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1324 {
1325         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1326         struct hugetlbfs_sb_info *sbinfo;
1327
1328         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1329         if (!sbinfo)
1330                 return -ENOMEM;
1331         sb->s_fs_info = sbinfo;
1332         spin_lock_init(&sbinfo->stat_lock);
1333         sbinfo->hstate          = ctx->hstate;
1334         sbinfo->max_inodes      = ctx->nr_inodes;
1335         sbinfo->free_inodes     = ctx->nr_inodes;
1336         sbinfo->spool           = NULL;
1337         sbinfo->uid             = ctx->uid;
1338         sbinfo->gid             = ctx->gid;
1339         sbinfo->mode            = ctx->mode;
1340
1341         /*
1342          * Allocate and initialize subpool if maximum or minimum size is
1343          * specified.  Any needed reservations (for minimum size) are taken
1344          * taken when the subpool is created.
1345          */
1346         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1347                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1348                                                      ctx->max_hpages,
1349                                                      ctx->min_hpages);
1350                 if (!sbinfo->spool)
1351                         goto out_free;
1352         }
1353         sb->s_maxbytes = MAX_LFS_FILESIZE;
1354         sb->s_blocksize = huge_page_size(ctx->hstate);
1355         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1356         sb->s_magic = HUGETLBFS_MAGIC;
1357         sb->s_op = &hugetlbfs_ops;
1358         sb->s_time_gran = 1;
1359
1360         /*
1361          * Due to the special and limited functionality of hugetlbfs, it does
1362          * not work well as a stacking filesystem.
1363          */
1364         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1365         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1366         if (!sb->s_root)
1367                 goto out_free;
1368         return 0;
1369 out_free:
1370         kfree(sbinfo->spool);
1371         kfree(sbinfo);
1372         return -ENOMEM;
1373 }
1374
1375 static int hugetlbfs_get_tree(struct fs_context *fc)
1376 {
1377         int err = hugetlbfs_validate(fc);
1378         if (err)
1379                 return err;
1380         return get_tree_nodev(fc, hugetlbfs_fill_super);
1381 }
1382
1383 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1384 {
1385         kfree(fc->fs_private);
1386 }
1387
1388 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1389         .free           = hugetlbfs_fs_context_free,
1390         .parse_param    = hugetlbfs_parse_param,
1391         .get_tree       = hugetlbfs_get_tree,
1392 };
1393
1394 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1395 {
1396         struct hugetlbfs_fs_context *ctx;
1397
1398         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1399         if (!ctx)
1400                 return -ENOMEM;
1401
1402         ctx->max_hpages = -1; /* No limit on size by default */
1403         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1404         ctx->uid        = current_fsuid();
1405         ctx->gid        = current_fsgid();
1406         ctx->mode       = 0755;
1407         ctx->hstate     = &default_hstate;
1408         ctx->min_hpages = -1; /* No default minimum size */
1409         ctx->max_val_type = NO_SIZE;
1410         ctx->min_val_type = NO_SIZE;
1411         fc->fs_private = ctx;
1412         fc->ops = &hugetlbfs_fs_context_ops;
1413         return 0;
1414 }
1415
1416 static struct file_system_type hugetlbfs_fs_type = {
1417         .name                   = "hugetlbfs",
1418         .init_fs_context        = hugetlbfs_init_fs_context,
1419         .parameters             = hugetlb_fs_parameters,
1420         .kill_sb                = kill_litter_super,
1421 };
1422
1423 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1424
1425 static int can_do_hugetlb_shm(void)
1426 {
1427         kgid_t shm_group;
1428         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1429         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1430 }
1431
1432 static int get_hstate_idx(int page_size_log)
1433 {
1434         struct hstate *h = hstate_sizelog(page_size_log);
1435
1436         if (!h)
1437                 return -1;
1438         return h - hstates;
1439 }
1440
1441 /*
1442  * Note that size should be aligned to proper hugepage size in caller side,
1443  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1444  */
1445 struct file *hugetlb_file_setup(const char *name, size_t size,
1446                                 vm_flags_t acctflag, struct user_struct **user,
1447                                 int creat_flags, int page_size_log)
1448 {
1449         struct inode *inode;
1450         struct vfsmount *mnt;
1451         int hstate_idx;
1452         struct file *file;
1453
1454         hstate_idx = get_hstate_idx(page_size_log);
1455         if (hstate_idx < 0)
1456                 return ERR_PTR(-ENODEV);
1457
1458         *user = NULL;
1459         mnt = hugetlbfs_vfsmount[hstate_idx];
1460         if (!mnt)
1461                 return ERR_PTR(-ENOENT);
1462
1463         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1464                 *user = current_user();
1465                 if (user_shm_lock(size, *user)) {
1466                         task_lock(current);
1467                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1468                                 current->comm, current->pid);
1469                         task_unlock(current);
1470                 } else {
1471                         *user = NULL;
1472                         return ERR_PTR(-EPERM);
1473                 }
1474         }
1475
1476         file = ERR_PTR(-ENOSPC);
1477         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1478         if (!inode)
1479                 goto out;
1480         if (creat_flags == HUGETLB_SHMFS_INODE)
1481                 inode->i_flags |= S_PRIVATE;
1482
1483         inode->i_size = size;
1484         clear_nlink(inode);
1485
1486         if (!hugetlb_reserve_pages(inode, 0,
1487                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1488                         acctflag))
1489                 file = ERR_PTR(-ENOMEM);
1490         else
1491                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1492                                         &hugetlbfs_file_operations);
1493         if (!IS_ERR(file))
1494                 return file;
1495
1496         iput(inode);
1497 out:
1498         if (*user) {
1499                 user_shm_unlock(size, *user);
1500                 *user = NULL;
1501         }
1502         return file;
1503 }
1504
1505 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1506 {
1507         struct fs_context *fc;
1508         struct vfsmount *mnt;
1509
1510         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1511         if (IS_ERR(fc)) {
1512                 mnt = ERR_CAST(fc);
1513         } else {
1514                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1515                 ctx->hstate = h;
1516                 mnt = fc_mount(fc);
1517                 put_fs_context(fc);
1518         }
1519         if (IS_ERR(mnt))
1520                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1521                        huge_page_size(h) >> 10);
1522         return mnt;
1523 }
1524
1525 static int __init init_hugetlbfs_fs(void)
1526 {
1527         struct vfsmount *mnt;
1528         struct hstate *h;
1529         int error;
1530         int i;
1531
1532         if (!hugepages_supported()) {
1533                 pr_info("disabling because there are no supported hugepage sizes\n");
1534                 return -ENOTSUPP;
1535         }
1536
1537         error = -ENOMEM;
1538         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1539                                         sizeof(struct hugetlbfs_inode_info),
1540                                         0, SLAB_ACCOUNT, init_once);
1541         if (hugetlbfs_inode_cachep == NULL)
1542                 goto out;
1543
1544         error = register_filesystem(&hugetlbfs_fs_type);
1545         if (error)
1546                 goto out_free;
1547
1548         /* default hstate mount is required */
1549         mnt = mount_one_hugetlbfs(&default_hstate);
1550         if (IS_ERR(mnt)) {
1551                 error = PTR_ERR(mnt);
1552                 goto out_unreg;
1553         }
1554         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1555
1556         /* other hstates are optional */
1557         i = 0;
1558         for_each_hstate(h) {
1559                 if (i == default_hstate_idx) {
1560                         i++;
1561                         continue;
1562                 }
1563
1564                 mnt = mount_one_hugetlbfs(h);
1565                 if (IS_ERR(mnt))
1566                         hugetlbfs_vfsmount[i] = NULL;
1567                 else
1568                         hugetlbfs_vfsmount[i] = mnt;
1569                 i++;
1570         }
1571
1572         return 0;
1573
1574  out_unreg:
1575         (void)unregister_filesystem(&hugetlbfs_fs_type);
1576  out_free:
1577         kmem_cache_destroy(hugetlbfs_inode_cachep);
1578  out:
1579         return error;
1580 }
1581 fs_initcall(init_hugetlbfs_fs)