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         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 = TASK_SIZE;
 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 = 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 = TASK_SIZE;
 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
 255         if (len & ~huge_page_mask(h))
 256                 return -EINVAL;
 257         if (len > TASK_SIZE)
 258                 return -ENOMEM;
 259
 260         if (flags & MAP_FIXED) {
 261                 if (prepare_hugepage_range(file, addr, len))
 262                         return -EINVAL;
 263                 return addr;
 264         }
 265
 266         if (addr) {
 267                 addr = ALIGN(addr, huge_page_size(h));
 268                 vma = find_vma(mm, addr);
 269                 if (TASK_SIZE - len >= addr &&
 270                     (!vma || addr + len <= vm_start_gap(vma)))
 271                         return addr;
 272         }
 273
 274         /*
 275          * Use mm->get_unmapped_area value as a hint to use topdown routine.
 276          * If architectures have special needs, they should define their own
 277          * version of hugetlb_get_unmapped_area.
 278          */
 279         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
 280                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 281                                 pgoff, flags);
 282         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 283                         pgoff, flags);
 284 }
 285 #endif
 286
 287 static size_t
 288 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 289                         struct iov_iter *to, unsigned long size)
 290 {
 291         size_t copied = 0;
 292         int i, chunksize;
 293
 294         /* Find which 4k chunk and offset with in that chunk */
 295         i = offset >> PAGE_SHIFT;
 296         offset = offset & ~PAGE_MASK;
 297
 298         while (size) {
 299                 size_t n;
 300                 chunksize = PAGE_SIZE;
 301                 if (offset)
 302                         chunksize -= offset;
 303                 if (chunksize > size)
 304                         chunksize = size;
 305                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 306                 copied += n;
 307                 if (n != chunksize)
 308                         return copied;
 309                 offset = 0;
 310                 size -= chunksize;
 311                 i++;
 312         }
 313         return copied;
 314 }
 315
 316 /*
 317  * Support for read() - Find the page attached to f_mapping and copy out the
 318  * data. Its *very* similar to generic_file_buffered_read(), we can't use that
 319  * since it has PAGE_SIZE assumptions.
 320  */
 321 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 322 {
 323         struct file *file = iocb->ki_filp;
 324         struct hstate *h = hstate_file(file);
 325         struct address_space *mapping = file->f_mapping;
 326         struct inode *inode = mapping->host;
 327         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 328         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 329         unsigned long end_index;
 330         loff_t isize;
 331         ssize_t retval = 0;
 332
 333         while (iov_iter_count(to)) {
 334                 struct page *page;
 335                 size_t nr, copied;
 336
 337                 /* nr is the maximum number of bytes to copy from this page */
 338                 nr = huge_page_size(h);
 339                 isize = i_size_read(inode);
 340                 if (!isize)
 341                         break;
 342                 end_index = (isize - 1) >> huge_page_shift(h);
 343                 if (index > end_index)
 344                         break;
 345                 if (index == end_index) {
 346                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 347                         if (nr <= offset)
 348                                 break;
 349                 }
 350                 nr = nr - offset;
 351
 352                 /* Find the page */
 353                 page = find_lock_page(mapping, index);
 354                 if (unlikely(page == NULL)) {
 355                         /*
 356                          * We have a HOLE, zero out the user-buffer for the
 357                          * length of the hole or request.
 358                          */
 359                         copied = iov_iter_zero(nr, to);
 360                 } else {
 361                         unlock_page(page);
 362
 363                         /*
 364                          * We have the page, copy it to user space buffer.
 365                          */
 366                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 367                         put_page(page);
 368                 }
 369                 offset += copied;
 370                 retval += copied;
 371                 if (copied != nr && iov_iter_count(to)) {
 372                         if (!retval)
 373                                 retval = -EFAULT;
 374                         break;
 375                 }
 376                 index += offset >> huge_page_shift(h);
 377                 offset &= ~huge_page_mask(h);
 378         }
 379         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 380         return retval;
 381 }
 382
 383 static int hugetlbfs_write_begin(struct file *file,
 384                         struct address_space *mapping,
 385                         loff_t pos, unsigned len, unsigned flags,
 386                         struct page **pagep, void **fsdata)
 387 {
 388         return -EINVAL;
 389 }
 390
 391 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 392                         loff_t pos, unsigned len, unsigned copied,
 393                         struct page *page, void *fsdata)
 394 {
 395         BUG();
 396         return -EINVAL;
 397 }
 398
 399 static void remove_huge_page(struct page *page)
 400 {
 401         ClearPageDirty(page);
 402         ClearPageUptodate(page);
 403         delete_from_page_cache(page);
 404 }
 405
 406 static void
 407 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
 408 {
 409         struct vm_area_struct *vma;
 410
 411         /*
 412          * end == 0 indicates that the entire range after
 413          * start should be unmapped.
 414          */
 415         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 416                 unsigned long v_offset;
 417                 unsigned long v_end;
 418
 419                 /*
 420                  * Can the expression below overflow on 32-bit arches?
 421                  * No, because the interval tree returns us only those vmas
 422                  * which overlap the truncated area starting at pgoff,
 423                  * and no vma on a 32-bit arch can span beyond the 4GB.
 424                  */
 425                 if (vma->vm_pgoff < start)
 426                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 427                 else
 428                         v_offset = 0;
 429
 430                 if (!end)
 431                         v_end = vma->vm_end;
 432                 else {
 433                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 434                                                         + vma->vm_start;
 435                         if (v_end > vma->vm_end)
 436                                 v_end = vma->vm_end;
 437                 }
 438
 439                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 440                                                                         NULL);
 441         }
 442 }
 443
 444 /*
 445  * remove_inode_hugepages handles two distinct cases: truncation and hole
 446  * punch.  There are subtle differences in operation for each case.
 447  *
 448  * truncation is indicated by end of range being LLONG_MAX
 449  *      In this case, we first scan the range and release found pages.
 450  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
 451  *      maps and global counts.  Page faults can not race with truncation
 452  *      in this routine.  hugetlb_no_page() holds i_mmap_rwsem and prevents
 453  *      page faults in the truncated range by checking i_size.  i_size is
 454  *      modified while holding i_mmap_rwsem.
 455  * hole punch is indicated if end is not LLONG_MAX
 456  *      In the hole punch case we scan the range and release found pages.
 457  *      Only when releasing a page is the associated region/reserve map
 458  *      deleted.  The region/reserve map for ranges without associated
 459  *      pages are not modified.  Page faults can race with hole punch.
 460  *      This is indicated if we find a mapped page.
 461  * Note: If the passed end of range value is beyond the end of file, but
 462  * not LLONG_MAX this routine still performs a hole punch operation.
 463  */
 464 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 465                                    loff_t lend)
 466 {
 467         struct hstate *h = hstate_inode(inode);
 468         struct address_space *mapping = &inode->i_data;
 469         const pgoff_t start = lstart >> huge_page_shift(h);
 470         const pgoff_t end = lend >> huge_page_shift(h);
 471         struct pagevec pvec;
 472         pgoff_t next, index;
 473         int i, freed = 0;
 474         bool truncate_op = (lend == LLONG_MAX);
 475
 476         pagevec_init(&pvec);
 477         next = start;
 478         while (next < end) {
 479                 /*
 480                  * When no more pages are found, we are done.
 481                  */
 482                 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
 483                         break;
 484
 485                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 486                         struct page *page = pvec.pages[i];
 487                         u32 hash = 0;
 488
 489                         index = page->index;
 490                         if (!truncate_op) {
 491                                 /*
 492                                  * Only need to hold the fault mutex in the
 493                                  * hole punch case.  This prevents races with
 494                                  * page faults.  Races are not possible in the
 495                                  * case of truncation.
 496                                  */
 497                                 hash = hugetlb_fault_mutex_hash(mapping, index);
 498                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 499                         }
 500
 501                         /*
 502                          * If page is mapped, it was faulted in after being
 503                          * unmapped in caller.  Unmap (again) now after taking
 504                          * the fault mutex.  The mutex will prevent faults
 505                          * until we finish removing the page.
 506                          *
 507                          * This race can only happen in the hole punch case.
 508                          * Getting here in a truncate operation is a bug.
 509                          */
 510                         if (unlikely(page_mapped(page))) {
 511                                 BUG_ON(truncate_op);
 512
 513                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 514                                 i_mmap_lock_write(mapping);
 515                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 516                                 hugetlb_vmdelete_list(&mapping->i_mmap,
 517                                         index * pages_per_huge_page(h),
 518                                         (index + 1) * pages_per_huge_page(h));
 519                                 i_mmap_unlock_write(mapping);
 520                         }
 521
 522                         lock_page(page);
 523                         /*
 524                          * We must free the huge page and remove from page
 525                          * cache (remove_huge_page) BEFORE removing the
 526                          * region/reserve map (hugetlb_unreserve_pages).  In
 527                          * rare out of memory conditions, removal of the
 528                          * region/reserve map could fail. Correspondingly,
 529                          * the subpool and global reserve usage count can need
 530                          * to be adjusted.
 531                          */
 532                         VM_BUG_ON(HPageRestoreReserve(page));
 533                         remove_huge_page(page);
 534                         freed++;
 535                         if (!truncate_op) {
 536                                 if (unlikely(hugetlb_unreserve_pages(inode,
 537                                                         index, index + 1, 1)))
 538                                         hugetlb_fix_reserve_counts(inode);
 539                         }
 540
 541                         unlock_page(page);
 542                         if (!truncate_op)
 543                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 544                 }
 545                 huge_pagevec_release(&pvec);
 546                 cond_resched();
 547         }
 548
 549         if (truncate_op)
 550                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 551 }
 552
 553 static void hugetlbfs_evict_inode(struct inode *inode)
 554 {
 555         struct resv_map *resv_map;
 556
 557         remove_inode_hugepages(inode, 0, LLONG_MAX);
 558
 559         /*
 560          * Get the resv_map from the address space embedded in the inode.
 561          * This is the address space which points to any resv_map allocated
 562          * at inode creation time.  If this is a device special inode,
 563          * i_mapping may not point to the original address space.
 564          */
 565         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
 566         /* Only regular and link inodes have associated reserve maps */
 567         if (resv_map)
 568                 resv_map_release(&resv_map->refs);
 569         clear_inode(inode);
 570 }
 571
 572 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 573 {
 574         pgoff_t pgoff;
 575         struct address_space *mapping = inode->i_mapping;
 576         struct hstate *h = hstate_inode(inode);
 577
 578         BUG_ON(offset & ~huge_page_mask(h));
 579         pgoff = offset >> PAGE_SHIFT;
 580
 581         i_mmap_lock_write(mapping);
 582         i_size_write(inode, offset);
 583         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 584                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 585         i_mmap_unlock_write(mapping);
 586         remove_inode_hugepages(inode, offset, LLONG_MAX);
 587 }
 588
 589 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 590 {
 591         struct hstate *h = hstate_inode(inode);
 592         loff_t hpage_size = huge_page_size(h);
 593         loff_t hole_start, hole_end;
 594
 595         /*
 596          * For hole punch round up the beginning offset of the hole and
 597          * round down the end.
 598          */
 599         hole_start = round_up(offset, hpage_size);
 600         hole_end = round_down(offset + len, hpage_size);
 601
 602         if (hole_end > hole_start) {
 603                 struct address_space *mapping = inode->i_mapping;
 604                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 605
 606                 inode_lock(inode);
 607
 608                 /* protected by i_rwsem */
 609                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 610                         inode_unlock(inode);
 611                         return -EPERM;
 612                 }
 613
 614                 i_mmap_lock_write(mapping);
 615                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 616                         hugetlb_vmdelete_list(&mapping->i_mmap,
 617                                                 hole_start >> PAGE_SHIFT,
 618                                                 hole_end  >> PAGE_SHIFT);
 619                 i_mmap_unlock_write(mapping);
 620                 remove_inode_hugepages(inode, hole_start, hole_end);
 621                 inode_unlock(inode);
 622         }
 623
 624         return 0;
 625 }
 626
 627 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 628                                 loff_t len)
 629 {
 630         struct inode *inode = file_inode(file);
 631         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 632         struct address_space *mapping = inode->i_mapping;
 633         struct hstate *h = hstate_inode(inode);
 634         struct vm_area_struct pseudo_vma;
 635         struct mm_struct *mm = current->mm;
 636         loff_t hpage_size = huge_page_size(h);
 637         unsigned long hpage_shift = huge_page_shift(h);
 638         pgoff_t start, index, end;
 639         int error;
 640         u32 hash;
 641
 642         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 643                 return -EOPNOTSUPP;
 644
 645         if (mode & FALLOC_FL_PUNCH_HOLE)
 646                 return hugetlbfs_punch_hole(inode, offset, len);
 647
 648         /*
 649          * Default preallocate case.
 650          * For this range, start is rounded down and end is rounded up
 651          * as well as being converted to page offsets.
 652          */
 653         start = offset >> hpage_shift;
 654         end = (offset + len + hpage_size - 1) >> hpage_shift;
 655
 656         inode_lock(inode);
 657
 658         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 659         error = inode_newsize_ok(inode, offset + len);
 660         if (error)
 661                 goto out;
 662
 663         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 664                 error = -EPERM;
 665                 goto out;
 666         }
 667
 668         /*
 669          * Initialize a pseudo vma as this is required by the huge page
 670          * allocation routines.  If NUMA is configured, use page index
 671          * as input to create an allocation policy.
 672          */
 673         vma_init(&pseudo_vma, mm);
 674         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 675         pseudo_vma.vm_file = file;
 676
 677         for (index = start; index < end; index++) {
 678                 /*
 679                  * This is supposed to be the vaddr where the page is being
 680                  * faulted in, but we have no vaddr here.
 681                  */
 682                 struct page *page;
 683                 unsigned long addr;
 684
 685                 cond_resched();
 686
 687                 /*
 688                  * fallocate(2) manpage permits EINTR; we may have been
 689                  * interrupted because we are using up too much memory.
 690                  */
 691                 if (signal_pending(current)) {
 692                         error = -EINTR;
 693                         break;
 694                 }
 695
 696                 /* Set numa allocation policy based on index */
 697                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 698
 699                 /* addr is the offset within the file (zero based) */
 700                 addr = index * hpage_size;
 701
 702                 /*
 703                  * fault mutex taken here, protects against fault path
 704                  * and hole punch.  inode_lock previously taken protects
 705                  * against truncation.
 706                  */
 707                 hash = hugetlb_fault_mutex_hash(mapping, index);
 708                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 709
 710                 /* See if already present in mapping to avoid alloc/free */
 711                 page = find_get_page(mapping, index);
 712                 if (page) {
 713                         put_page(page);
 714                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 715                         hugetlb_drop_vma_policy(&pseudo_vma);
 716                         continue;
 717                 }
 718
 719                 /*
 720                  * Allocate page without setting the avoid_reserve argument.
 721                  * There certainly are no reserves associated with the
 722                  * pseudo_vma.  However, there could be shared mappings with
 723                  * reserves for the file at the inode level.  If we fallocate
 724                  * pages in these areas, we need to consume the reserves
 725                  * to keep reservation accounting consistent.
 726                  */
 727                 page = alloc_huge_page(&pseudo_vma, addr, 0);
 728                 hugetlb_drop_vma_policy(&pseudo_vma);
 729                 if (IS_ERR(page)) {
 730                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 731                         error = PTR_ERR(page);
 732                         goto out;
 733                 }
 734                 clear_huge_page(page, addr, pages_per_huge_page(h));
 735                 __SetPageUptodate(page);
 736                 error = huge_add_to_page_cache(page, mapping, index);
 737                 if (unlikely(error)) {
 738                         put_page(page);
 739                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 740                         goto out;
 741                 }
 742
 743                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 744
 745                 SetHPageMigratable(page);
 746                 /*
 747                  * unlock_page because locked by add_to_page_cache()
 748                  * put_page() due to reference from alloc_huge_page()
 749                  */
 750                 unlock_page(page);
 751                 put_page(page);
 752         }
 753
 754         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 755                 i_size_write(inode, offset + len);
 756         inode->i_ctime = current_time(inode);
 757 out:
 758         inode_unlock(inode);
 759         return error;
 760 }
 761
 762 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
 763                              struct dentry *dentry, struct iattr *attr)
 764 {
 765         struct inode *inode = d_inode(dentry);
 766         struct hstate *h = hstate_inode(inode);
 767         int error;
 768         unsigned int ia_valid = attr->ia_valid;
 769         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 770
 771         error = setattr_prepare(&init_user_ns, dentry, attr);
 772         if (error)
 773                 return error;
 774
 775         if (ia_valid & ATTR_SIZE) {
 776                 loff_t oldsize = inode->i_size;
 777                 loff_t newsize = attr->ia_size;
 778
 779                 if (newsize & ~huge_page_mask(h))
 780                         return -EINVAL;
 781                 /* protected by i_rwsem */
 782                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 783                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 784                         return -EPERM;
 785                 hugetlb_vmtruncate(inode, newsize);
 786         }
 787
 788         setattr_copy(&init_user_ns, inode, attr);
 789         mark_inode_dirty(inode);
 790         return 0;
 791 }
 792
 793 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 794                                         struct hugetlbfs_fs_context *ctx)
 795 {
 796         struct inode *inode;
 797
 798         inode = new_inode(sb);
 799         if (inode) {
 800                 inode->i_ino = get_next_ino();
 801                 inode->i_mode = S_IFDIR | ctx->mode;
 802                 inode->i_uid = ctx->uid;
 803                 inode->i_gid = ctx->gid;
 804                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 805                 inode->i_op = &hugetlbfs_dir_inode_operations;
 806                 inode->i_fop = &simple_dir_operations;
 807                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 808                 inc_nlink(inode);
 809                 lockdep_annotate_inode_mutex_key(inode);
 810         }
 811         return inode;
 812 }
 813
 814 /*
 815  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 816  * be taken from reclaim -- unlike regular filesystems. This needs an
 817  * annotation because huge_pmd_share() does an allocation under hugetlb's
 818  * i_mmap_rwsem.
 819  */
 820 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 821
 822 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 823                                         struct inode *dir,
 824                                         umode_t mode, dev_t dev)
 825 {
 826         struct inode *inode;
 827         struct resv_map *resv_map = NULL;
 828
 829         /*
 830          * Reserve maps are only needed for inodes that can have associated
 831          * page allocations.
 832          */
 833         if (S_ISREG(mode) || S_ISLNK(mode)) {
 834                 resv_map = resv_map_alloc();
 835                 if (!resv_map)
 836                         return NULL;
 837         }
 838
 839         inode = new_inode(sb);
 840         if (inode) {
 841                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 842
 843                 inode->i_ino = get_next_ino();
 844                 inode_init_owner(&init_user_ns, inode, dir, mode);
 845                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 846                                 &hugetlbfs_i_mmap_rwsem_key);
 847                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 848                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 849                 inode->i_mapping->private_data = resv_map;
 850                 info->seals = F_SEAL_SEAL;
 851                 switch (mode & S_IFMT) {
 852                 default:
 853                         init_special_inode(inode, mode, dev);
 854                         break;
 855                 case S_IFREG:
 856                         inode->i_op = &hugetlbfs_inode_operations;
 857                         inode->i_fop = &hugetlbfs_file_operations;
 858                         break;
 859                 case S_IFDIR:
 860                         inode->i_op = &hugetlbfs_dir_inode_operations;
 861                         inode->i_fop = &simple_dir_operations;
 862
 863                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 864                         inc_nlink(inode);
 865                         break;
 866                 case S_IFLNK:
 867                         inode->i_op = &page_symlink_inode_operations;
 868                         inode_nohighmem(inode);
 869                         break;
 870                 }
 871                 lockdep_annotate_inode_mutex_key(inode);
 872         } else {
 873                 if (resv_map)
 874                         kref_put(&resv_map->refs, resv_map_release);
 875         }
 876
 877         return inode;
 878 }
 879
 880 /*
 881  * File creation. Allocate an inode, and we're done..
 882  */
 883 static int do_hugetlbfs_mknod(struct inode *dir,
 884                         struct dentry *dentry,
 885                         umode_t mode,
 886                         dev_t dev,
 887                         bool tmpfile)
 888 {
 889         struct inode *inode;
 890         int error = -ENOSPC;
 891
 892         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 893         if (inode) {
 894                 dir->i_ctime = dir->i_mtime = current_time(dir);
 895                 if (tmpfile) {
 896                         d_tmpfile(dentry, inode);
 897                 } else {
 898                         d_instantiate(dentry, inode);
 899                         dget(dentry);/* Extra count - pin the dentry in core */
 900                 }
 901                 error = 0;
 902         }
 903         return error;
 904 }
 905
 906 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
 907                            struct dentry *dentry, umode_t mode, dev_t dev)
 908 {
 909         return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
 910 }
 911
 912 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
 913                            struct dentry *dentry, umode_t mode)
 914 {
 915         int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
 916                                      mode | S_IFDIR, 0);
 917         if (!retval)
 918                 inc_nlink(dir);
 919         return retval;
 920 }
 921
 922 static int hugetlbfs_create(struct user_namespace *mnt_userns,
 923                             struct inode *dir, struct dentry *dentry,
 924                             umode_t mode, bool excl)
 925 {
 926         return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
 927 }
 928
 929 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
 930                              struct inode *dir, struct dentry *dentry,
 931                              umode_t mode)
 932 {
 933         return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
 934 }
 935
 936 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
 937                              struct inode *dir, struct dentry *dentry,
 938                              const char *symname)
 939 {
 940         struct inode *inode;
 941         int error = -ENOSPC;
 942
 943         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 944         if (inode) {
 945                 int l = strlen(symname)+1;
 946                 error = page_symlink(inode, symname, l);
 947                 if (!error) {
 948                         d_instantiate(dentry, inode);
 949                         dget(dentry);
 950                 } else
 951                         iput(inode);
 952         }
 953         dir->i_ctime = dir->i_mtime = current_time(dir);
 954
 955         return error;
 956 }
 957
 958 static int hugetlbfs_migrate_page(struct address_space *mapping,
 959                                 struct page *newpage, struct page *page,
 960                                 enum migrate_mode mode)
 961 {
 962         int rc;
 963
 964         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 965         if (rc != MIGRATEPAGE_SUCCESS)
 966                 return rc;
 967
 968         if (hugetlb_page_subpool(page)) {
 969                 hugetlb_set_page_subpool(newpage, hugetlb_page_subpool(page));
 970                 hugetlb_set_page_subpool(page, NULL);
 971         }
 972
 973         if (mode != MIGRATE_SYNC_NO_COPY)
 974                 migrate_page_copy(newpage, page);
 975         else
 976                 migrate_page_states(newpage, page);
 977
 978         return MIGRATEPAGE_SUCCESS;
 979 }
 980
 981 static int hugetlbfs_error_remove_page(struct address_space *mapping,
 982                                 struct page *page)
 983 {
 984         struct inode *inode = mapping->host;
 985         pgoff_t index = page->index;
 986
 987         remove_huge_page(page);
 988         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
 989                 hugetlb_fix_reserve_counts(inode);
 990
 991         return 0;
 992 }
 993
 994 /*
 995  * Display the mount options in /proc/mounts.
 996  */
 997 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
 998 {
 999         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1000         struct hugepage_subpool *spool = sbinfo->spool;
1001         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1002         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1003         char mod;
1004
1005         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1006                 seq_printf(m, ",uid=%u",
1007                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1008         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1009                 seq_printf(m, ",gid=%u",
1010                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1011         if (sbinfo->mode != 0755)
1012                 seq_printf(m, ",mode=%o", sbinfo->mode);
1013         if (sbinfo->max_inodes != -1)
1014                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1015
1016         hpage_size /= 1024;
1017         mod = 'K';
1018         if (hpage_size >= 1024) {
1019                 hpage_size /= 1024;
1020                 mod = 'M';
1021         }
1022         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1023         if (spool) {
1024                 if (spool->max_hpages != -1)
1025                         seq_printf(m, ",size=%llu",
1026                                    (unsigned long long)spool->max_hpages << hpage_shift);
1027                 if (spool->min_hpages != -1)
1028                         seq_printf(m, ",min_size=%llu",
1029                                    (unsigned long long)spool->min_hpages << hpage_shift);
1030         }
1031         return 0;
1032 }
1033
1034 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1035 {
1036         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1037         struct hstate *h = hstate_inode(d_inode(dentry));
1038
1039         buf->f_type = HUGETLBFS_MAGIC;
1040         buf->f_bsize = huge_page_size(h);
1041         if (sbinfo) {
1042                 spin_lock(&sbinfo->stat_lock);
1043                 /* If no limits set, just report 0 for max/free/used
1044                  * blocks, like simple_statfs() */
1045                 if (sbinfo->spool) {
1046                         long free_pages;
1047
1048                         spin_lock(&sbinfo->spool->lock);
1049                         buf->f_blocks = sbinfo->spool->max_hpages;
1050                         free_pages = sbinfo->spool->max_hpages
1051                                 - sbinfo->spool->used_hpages;
1052                         buf->f_bavail = buf->f_bfree = free_pages;
1053                         spin_unlock(&sbinfo->spool->lock);
1054                         buf->f_files = sbinfo->max_inodes;
1055                         buf->f_ffree = sbinfo->free_inodes;
1056                 }
1057                 spin_unlock(&sbinfo->stat_lock);
1058         }
1059         buf->f_namelen = NAME_MAX;
1060         return 0;
1061 }
1062
1063 static void hugetlbfs_put_super(struct super_block *sb)
1064 {
1065         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1066
1067         if (sbi) {
1068                 sb->s_fs_info = NULL;
1069
1070                 if (sbi->spool)
1071                         hugepage_put_subpool(sbi->spool);
1072
1073                 kfree(sbi);
1074         }
1075 }
1076
1077 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1078 {
1079         if (sbinfo->free_inodes >= 0) {
1080                 spin_lock(&sbinfo->stat_lock);
1081                 if (unlikely(!sbinfo->free_inodes)) {
1082                         spin_unlock(&sbinfo->stat_lock);
1083                         return 0;
1084                 }
1085                 sbinfo->free_inodes--;
1086                 spin_unlock(&sbinfo->stat_lock);
1087         }
1088
1089         return 1;
1090 }
1091
1092 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1093 {
1094         if (sbinfo->free_inodes >= 0) {
1095                 spin_lock(&sbinfo->stat_lock);
1096                 sbinfo->free_inodes++;
1097                 spin_unlock(&sbinfo->stat_lock);
1098         }
1099 }
1100
1101
1102 static struct kmem_cache *hugetlbfs_inode_cachep;
1103
1104 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1105 {
1106         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1107         struct hugetlbfs_inode_info *p;
1108
1109         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1110                 return NULL;
1111         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1112         if (unlikely(!p)) {
1113                 hugetlbfs_inc_free_inodes(sbinfo);
1114                 return NULL;
1115         }
1116
1117         /*
1118          * Any time after allocation, hugetlbfs_destroy_inode can be called
1119          * for the inode.  mpol_free_shared_policy is unconditionally called
1120          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1121          * in case of a quick call to destroy.
1122          *
1123          * Note that the policy is initialized even if we are creating a
1124          * private inode.  This simplifies hugetlbfs_destroy_inode.
1125          */
1126         mpol_shared_policy_init(&p->policy, NULL);
1127
1128         return &p->vfs_inode;
1129 }
1130
1131 static void hugetlbfs_free_inode(struct inode *inode)
1132 {
1133         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1134 }
1135
1136 static void hugetlbfs_destroy_inode(struct inode *inode)
1137 {
1138         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1139         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1140 }
1141
1142 static const struct address_space_operations hugetlbfs_aops = {
1143         .write_begin    = hugetlbfs_write_begin,
1144         .write_end      = hugetlbfs_write_end,
1145         .set_page_dirty =  __set_page_dirty_no_writeback,
1146         .migratepage    = hugetlbfs_migrate_page,
1147         .error_remove_page      = hugetlbfs_error_remove_page,
1148 };
1149
1150
1151 static void init_once(void *foo)
1152 {
1153         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1154
1155         inode_init_once(&ei->vfs_inode);
1156 }
1157
1158 const struct file_operations hugetlbfs_file_operations = {
1159         .read_iter              = hugetlbfs_read_iter,
1160         .mmap                   = hugetlbfs_file_mmap,
1161         .fsync                  = noop_fsync,
1162         .get_unmapped_area      = hugetlb_get_unmapped_area,
1163         .llseek                 = default_llseek,
1164         .fallocate              = hugetlbfs_fallocate,
1165 };
1166
1167 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1168         .create         = hugetlbfs_create,
1169         .lookup         = simple_lookup,
1170         .link           = simple_link,
1171         .unlink         = simple_unlink,
1172         .symlink        = hugetlbfs_symlink,
1173         .mkdir          = hugetlbfs_mkdir,
1174         .rmdir          = simple_rmdir,
1175         .mknod          = hugetlbfs_mknod,
1176         .rename         = simple_rename,
1177         .setattr        = hugetlbfs_setattr,
1178         .tmpfile        = hugetlbfs_tmpfile,
1179 };
1180
1181 static const struct inode_operations hugetlbfs_inode_operations = {
1182         .setattr        = hugetlbfs_setattr,
1183 };
1184
1185 static const struct super_operations hugetlbfs_ops = {
1186         .alloc_inode    = hugetlbfs_alloc_inode,
1187         .free_inode     = hugetlbfs_free_inode,
1188         .destroy_inode  = hugetlbfs_destroy_inode,
1189         .evict_inode    = hugetlbfs_evict_inode,
1190         .statfs         = hugetlbfs_statfs,
1191         .put_super      = hugetlbfs_put_super,
1192         .show_options   = hugetlbfs_show_options,
1193 };
1194
1195 /*
1196  * Convert size option passed from command line to number of huge pages
1197  * in the pool specified by hstate.  Size option could be in bytes
1198  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1199  */
1200 static long
1201 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1202                          enum hugetlbfs_size_type val_type)
1203 {
1204         if (val_type == NO_SIZE)
1205                 return -1;
1206
1207         if (val_type == SIZE_PERCENT) {
1208                 size_opt <<= huge_page_shift(h);
1209                 size_opt *= h->max_huge_pages;
1210                 do_div(size_opt, 100);
1211         }
1212
1213         size_opt >>= huge_page_shift(h);
1214         return size_opt;
1215 }
1216
1217 /*
1218  * Parse one mount parameter.
1219  */
1220 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1221 {
1222         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1223         struct fs_parse_result result;
1224         char *rest;
1225         unsigned long ps;
1226         int opt;
1227
1228         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1229         if (opt < 0)
1230                 return opt;
1231
1232         switch (opt) {
1233         case Opt_uid:
1234                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1235                 if (!uid_valid(ctx->uid))
1236                         goto bad_val;
1237                 return 0;
1238
1239         case Opt_gid:
1240                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1241                 if (!gid_valid(ctx->gid))
1242                         goto bad_val;
1243                 return 0;
1244
1245         case Opt_mode:
1246                 ctx->mode = result.uint_32 & 01777U;
1247                 return 0;
1248
1249         case Opt_size:
1250                 /* memparse() will accept a K/M/G without a digit */
1251                 if (!isdigit(param->string[0]))
1252                         goto bad_val;
1253                 ctx->max_size_opt = memparse(param->string, &rest);
1254                 ctx->max_val_type = SIZE_STD;
1255                 if (*rest == '%')
1256                         ctx->max_val_type = SIZE_PERCENT;
1257                 return 0;
1258
1259         case Opt_nr_inodes:
1260                 /* memparse() will accept a K/M/G without a digit */
1261                 if (!isdigit(param->string[0]))
1262                         goto bad_val;
1263                 ctx->nr_inodes = memparse(param->string, &rest);
1264                 return 0;
1265
1266         case Opt_pagesize:
1267                 ps = memparse(param->string, &rest);
1268                 ctx->hstate = size_to_hstate(ps);
1269                 if (!ctx->hstate) {
1270                         pr_err("Unsupported page size %lu MB\n", ps >> 20);
1271                         return -EINVAL;
1272                 }
1273                 return 0;
1274
1275         case Opt_min_size:
1276                 /* memparse() will accept a K/M/G without a digit */
1277                 if (!isdigit(param->string[0]))
1278                         goto bad_val;
1279                 ctx->min_size_opt = memparse(param->string, &rest);
1280                 ctx->min_val_type = SIZE_STD;
1281                 if (*rest == '%')
1282                         ctx->min_val_type = SIZE_PERCENT;
1283                 return 0;
1284
1285         default:
1286                 return -EINVAL;
1287         }
1288
1289 bad_val:
1290         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1291                       param->string, param->key);
1292 }
1293
1294 /*
1295  * Validate the parsed options.
1296  */
1297 static int hugetlbfs_validate(struct fs_context *fc)
1298 {
1299         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1300
1301         /*
1302          * Use huge page pool size (in hstate) to convert the size
1303          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1304          */
1305         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1306                                                    ctx->max_size_opt,
1307                                                    ctx->max_val_type);
1308         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1309                                                    ctx->min_size_opt,
1310                                                    ctx->min_val_type);
1311
1312         /*
1313          * If max_size was specified, then min_size must be smaller
1314          */
1315         if (ctx->max_val_type > NO_SIZE &&
1316             ctx->min_hpages > ctx->max_hpages) {
1317                 pr_err("Minimum size can not be greater than maximum size\n");
1318                 return -EINVAL;
1319         }
1320
1321         return 0;
1322 }
1323
1324 static int
1325 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1326 {
1327         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1328         struct hugetlbfs_sb_info *sbinfo;
1329
1330         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1331         if (!sbinfo)
1332                 return -ENOMEM;
1333         sb->s_fs_info = sbinfo;
1334         spin_lock_init(&sbinfo->stat_lock);
1335         sbinfo->hstate          = ctx->hstate;
1336         sbinfo->max_inodes      = ctx->nr_inodes;
1337         sbinfo->free_inodes     = ctx->nr_inodes;
1338         sbinfo->spool           = NULL;
1339         sbinfo->uid             = ctx->uid;
1340         sbinfo->gid             = ctx->gid;
1341         sbinfo->mode            = ctx->mode;
1342
1343         /*
1344          * Allocate and initialize subpool if maximum or minimum size is
1345          * specified.  Any needed reservations (for minimum size) are taken
1346          * taken when the subpool is created.
1347          */
1348         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1349                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1350                                                      ctx->max_hpages,
1351                                                      ctx->min_hpages);
1352                 if (!sbinfo->spool)
1353                         goto out_free;
1354         }
1355         sb->s_maxbytes = MAX_LFS_FILESIZE;
1356         sb->s_blocksize = huge_page_size(ctx->hstate);
1357         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1358         sb->s_magic = HUGETLBFS_MAGIC;
1359         sb->s_op = &hugetlbfs_ops;
1360         sb->s_time_gran = 1;
1361
1362         /*
1363          * Due to the special and limited functionality of hugetlbfs, it does
1364          * not work well as a stacking filesystem.
1365          */
1366         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1367         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1368         if (!sb->s_root)
1369                 goto out_free;
1370         return 0;
1371 out_free:
1372         kfree(sbinfo->spool);
1373         kfree(sbinfo);
1374         return -ENOMEM;
1375 }
1376
1377 static int hugetlbfs_get_tree(struct fs_context *fc)
1378 {
1379         int err = hugetlbfs_validate(fc);
1380         if (err)
1381                 return err;
1382         return get_tree_nodev(fc, hugetlbfs_fill_super);
1383 }
1384
1385 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1386 {
1387         kfree(fc->fs_private);
1388 }
1389
1390 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1391         .free           = hugetlbfs_fs_context_free,
1392         .parse_param    = hugetlbfs_parse_param,
1393         .get_tree       = hugetlbfs_get_tree,
1394 };
1395
1396 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1397 {
1398         struct hugetlbfs_fs_context *ctx;
1399
1400         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1401         if (!ctx)
1402                 return -ENOMEM;
1403
1404         ctx->max_hpages = -1; /* No limit on size by default */
1405         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1406         ctx->uid        = current_fsuid();
1407         ctx->gid        = current_fsgid();
1408         ctx->mode       = 0755;
1409         ctx->hstate     = &default_hstate;
1410         ctx->min_hpages = -1; /* No default minimum size */
1411         ctx->max_val_type = NO_SIZE;
1412         ctx->min_val_type = NO_SIZE;
1413         fc->fs_private = ctx;
1414         fc->ops = &hugetlbfs_fs_context_ops;
1415         return 0;
1416 }
1417
1418 static struct file_system_type hugetlbfs_fs_type = {
1419         .name                   = "hugetlbfs",
1420         .init_fs_context        = hugetlbfs_init_fs_context,
1421         .parameters             = hugetlb_fs_parameters,
1422         .kill_sb                = kill_litter_super,
1423 };
1424
1425 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1426
1427 static int can_do_hugetlb_shm(void)
1428 {
1429         kgid_t shm_group;
1430         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1431         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1432 }
1433
1434 static int get_hstate_idx(int page_size_log)
1435 {
1436         struct hstate *h = hstate_sizelog(page_size_log);
1437
1438         if (!h)
1439                 return -1;
1440         return hstate_index(h);
1441 }
1442
1443 /*
1444  * Note that size should be aligned to proper hugepage size in caller side,
1445  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1446  */
1447 struct file *hugetlb_file_setup(const char *name, size_t size,
1448                                 vm_flags_t acctflag, struct user_struct **user,
1449                                 int creat_flags, int page_size_log)
1450 {
1451         struct inode *inode;
1452         struct vfsmount *mnt;
1453         int hstate_idx;
1454         struct file *file;
1455
1456         hstate_idx = get_hstate_idx(page_size_log);
1457         if (hstate_idx < 0)
1458                 return ERR_PTR(-ENODEV);
1459
1460         *user = NULL;
1461         mnt = hugetlbfs_vfsmount[hstate_idx];
1462         if (!mnt)
1463                 return ERR_PTR(-ENOENT);
1464
1465         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1466                 *user = current_user();
1467                 if (user_shm_lock(size, *user)) {
1468                         task_lock(current);
1469                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1470                                 current->comm, current->pid);
1471                         task_unlock(current);
1472                 } else {
1473                         *user = NULL;
1474                         return ERR_PTR(-EPERM);
1475                 }
1476         }
1477
1478         file = ERR_PTR(-ENOSPC);
1479         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1480         if (!inode)
1481                 goto out;
1482         if (creat_flags == HUGETLB_SHMFS_INODE)
1483                 inode->i_flags |= S_PRIVATE;
1484
1485         inode->i_size = size;
1486         clear_nlink(inode);
1487
1488         if (!hugetlb_reserve_pages(inode, 0,
1489                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1490                         acctflag))
1491                 file = ERR_PTR(-ENOMEM);
1492         else
1493                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1494                                         &hugetlbfs_file_operations);
1495         if (!IS_ERR(file))
1496                 return file;
1497
1498         iput(inode);
1499 out:
1500         if (*user) {
1501                 user_shm_unlock(size, *user);
1502                 *user = NULL;
1503         }
1504         return file;
1505 }
1506
1507 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1508 {
1509         struct fs_context *fc;
1510         struct vfsmount *mnt;
1511
1512         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1513         if (IS_ERR(fc)) {
1514                 mnt = ERR_CAST(fc);
1515         } else {
1516                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1517                 ctx->hstate = h;
1518                 mnt = fc_mount(fc);
1519                 put_fs_context(fc);
1520         }
1521         if (IS_ERR(mnt))
1522                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1523                        huge_page_size(h) >> 10);
1524         return mnt;
1525 }
1526
1527 static int __init init_hugetlbfs_fs(void)
1528 {
1529         struct vfsmount *mnt;
1530         struct hstate *h;
1531         int error;
1532         int i;
1533
1534         if (!hugepages_supported()) {
1535                 pr_info("disabling because there are no supported hugepage sizes\n");
1536                 return -ENOTSUPP;
1537         }
1538
1539         error = -ENOMEM;
1540         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1541                                         sizeof(struct hugetlbfs_inode_info),
1542                                         0, SLAB_ACCOUNT, init_once);
1543         if (hugetlbfs_inode_cachep == NULL)
1544                 goto out;
1545
1546         error = register_filesystem(&hugetlbfs_fs_type);
1547         if (error)
1548                 goto out_free;
1549
1550         /* default hstate mount is required */
1551         mnt = mount_one_hugetlbfs(&default_hstate);
1552         if (IS_ERR(mnt)) {
1553                 error = PTR_ERR(mnt);
1554                 goto out_unreg;
1555         }
1556         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1557
1558         /* other hstates are optional */
1559         i = 0;
1560         for_each_hstate(h) {
1561                 if (i == default_hstate_idx) {
1562                         i++;
1563                         continue;
1564                 }
1565
1566                 mnt = mount_one_hugetlbfs(h);
1567                 if (IS_ERR(mnt))
1568                         hugetlbfs_vfsmount[i] = NULL;
1569                 else
1570                         hugetlbfs_vfsmount[i] = mnt;
1571                 i++;
1572         }
1573
1574         return 0;
1575
1576  out_unreg:
1577         (void)unregister_filesystem(&hugetlbfs_fs_type);
1578  out_free:
1579         kmem_cache_destroy(hugetlbfs_inode_cachep);
1580  out:
1581         return error;
1582 }
1583 fs_initcall(init_hugetlbfs_fs)