fs/hugetlbfs/inode.c

   1 /*
   2  * hugetlbpage-backed filesystem.  Based on ramfs.
   3  *
   4  * Nadia Yvette Chambers, 2002
   5  *
   6  * Copyright (C) 2002 Linus Torvalds.
   7  * License: GPL
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/thread_info.h>
  13 #include <asm/current.h>
  14 #include <linux/sched/signal.h>         /* remove ASAP */
  15 #include <linux/falloc.h>
  16 #include <linux/fs.h>
  17 #include <linux/mount.h>
  18 #include <linux/file.h>
  19 #include <linux/kernel.h>
  20 #include <linux/writeback.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/highmem.h>
  23 #include <linux/init.h>
  24 #include <linux/string.h>
  25 #include <linux/capability.h>
  26 #include <linux/ctype.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/hugetlb.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/fs_parser.h>
  31 #include <linux/mman.h>
  32 #include <linux/slab.h>
  33 #include <linux/dnotify.h>
  34 #include <linux/statfs.h>
  35 #include <linux/security.h>
  36 #include <linux/magic.h>
  37 #include <linux/migrate.h>
  38 #include <linux/uio.h>
  39
  40 #include <linux/uaccess.h>
  41 #include <linux/sched/mm.h>
  42
  43 static const struct super_operations hugetlbfs_ops;
  44 static const struct address_space_operations hugetlbfs_aops;
  45 const struct file_operations hugetlbfs_file_operations;
  46 static const struct inode_operations hugetlbfs_dir_inode_operations;
  47 static const struct inode_operations hugetlbfs_inode_operations;
  48
  49 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
  50
  51 struct hugetlbfs_fs_context {
  52         struct hstate           *hstate;
  53         unsigned long long      max_size_opt;
  54         unsigned long long      min_size_opt;
  55         long                    max_hpages;
  56         long                    nr_inodes;
  57         long                    min_hpages;
  58         enum hugetlbfs_size_type max_val_type;
  59         enum hugetlbfs_size_type min_val_type;
  60         kuid_t                  uid;
  61         kgid_t                  gid;
  62         umode_t                 mode;
  63 };
  64
  65 int sysctl_hugetlb_shm_group;
  66
  67 enum hugetlb_param {
  68         Opt_gid,
  69         Opt_min_size,
  70         Opt_mode,
  71         Opt_nr_inodes,
  72         Opt_pagesize,
  73         Opt_size,
  74         Opt_uid,
  75 };
  76
  77 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
  78         fsparam_u32   ("gid",           Opt_gid),
  79         fsparam_string("min_size",      Opt_min_size),
  80         fsparam_u32oct("mode",          Opt_mode),
  81         fsparam_string("nr_inodes",     Opt_nr_inodes),
  82         fsparam_string("pagesize",      Opt_pagesize),
  83         fsparam_string("size",          Opt_size),
  84         fsparam_u32   ("uid",           Opt_uid),
  85         {}
  86 };
  87
  88 #ifdef CONFIG_NUMA
  89 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  90                                         struct inode *inode, pgoff_t index)
  91 {
  92         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  93                                                         index);
  94 }
  95
  96 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  97 {
  98         mpol_cond_put(vma->vm_policy);
  99 }
 100 #else
 101 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
 102                                         struct inode *inode, pgoff_t index)
 103 {
 104 }
 105
 106 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
 107 {
 108 }
 109 #endif
 110
 111 static void huge_pagevec_release(struct pagevec *pvec)
 112 {
 113         int i;
 114
 115         for (i = 0; i < pagevec_count(pvec); ++i)
 116                 put_page(pvec->pages[i]);
 117
 118         pagevec_reinit(pvec);
 119 }
 120
 121 /*
 122  * Mask used when checking the page offset value passed in via system
 123  * calls.  This value will be converted to a loff_t which is signed.
 124  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 125  * value.  The extra bit (- 1 in the shift value) is to take the sign
 126  * bit into account.
 127  */
 128 #define PGOFF_LOFFT_MAX \
 129         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 130
 131 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 132 {
 133         struct inode *inode = file_inode(file);
 134         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 135         loff_t len, vma_len;
 136         int ret;
 137         struct hstate *h = hstate_file(file);
 138
 139         /*
 140          * vma address alignment (but not the pgoff alignment) has
 141          * already been checked by prepare_hugepage_range.  If you add
 142          * any error returns here, do so after setting VM_HUGETLB, so
 143          * is_vm_hugetlb_page tests below unmap_region go the right
 144          * way when do_mmap unwinds (may be important on powerpc
 145          * and ia64).
 146          */
 147         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 148         vma->vm_ops = &hugetlb_vm_ops;
 149
 150         ret = seal_check_future_write(info->seals, vma);
 151         if (ret)
 152                 return ret;
 153
 154         /*
 155          * page based offset in vm_pgoff could be sufficiently large to
 156          * overflow a loff_t when converted to byte offset.  This can
 157          * only happen on architectures where sizeof(loff_t) ==
 158          * sizeof(unsigned long).  So, only check in those instances.
 159          */
 160         if (sizeof(unsigned long) == sizeof(loff_t)) {
 161                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 162                         return -EINVAL;
 163         }
 164
 165         /* must be huge page aligned */
 166         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 167                 return -EINVAL;
 168
 169         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 170         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 171         /* check for overflow */
 172         if (len < vma_len)
 173                 return -EINVAL;
 174
 175         inode_lock(inode);
 176         file_accessed(file);
 177
 178         ret = -ENOMEM;
 179         if (!hugetlb_reserve_pages(inode,
 180                                 vma->vm_pgoff >> huge_page_order(h),
 181                                 len >> huge_page_shift(h), vma,
 182                                 vma->vm_flags))
 183                 goto out;
 184
 185         ret = 0;
 186         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 187                 i_size_write(inode, len);
 188 out:
 189         inode_unlock(inode);
 190
 191         return ret;
 192 }
 193
 194 /*
 195  * Called under mmap_write_lock(mm).
 196  */
 197
 198 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 199 static unsigned long
 200 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
 201                 unsigned long len, unsigned long pgoff, unsigned long flags)
 202 {
 203         struct hstate *h = hstate_file(file);
 204         struct vm_unmapped_area_info info;
 205
 206         info.flags = 0;
 207         info.length = len;
 208         info.low_limit = current->mm->mmap_base;
 209         info.high_limit = arch_get_mmap_end(addr);
 210         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 211         info.align_offset = 0;
 212         return vm_unmapped_area(&info);
 213 }
 214
 215 static unsigned long
 216 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
 217                 unsigned long len, unsigned long pgoff, unsigned long flags)
 218 {
 219         struct hstate *h = hstate_file(file);
 220         struct vm_unmapped_area_info info;
 221
 222         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 223         info.length = len;
 224         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
 225         info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
 226         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 227         info.align_offset = 0;
 228         addr = vm_unmapped_area(&info);
 229
 230         /*
 231          * A failed mmap() very likely causes application failure,
 232          * so fall back to the bottom-up function here. This scenario
 233          * can happen with large stack limits and large mmap()
 234          * allocations.
 235          */
 236         if (unlikely(offset_in_page(addr))) {
 237                 VM_BUG_ON(addr != -ENOMEM);
 238                 info.flags = 0;
 239                 info.low_limit = current->mm->mmap_base;
 240                 info.high_limit = arch_get_mmap_end(addr);
 241                 addr = vm_unmapped_area(&info);
 242         }
 243
 244         return addr;
 245 }
 246
 247 static unsigned long
 248 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 249                 unsigned long len, unsigned long pgoff, unsigned long flags)
 250 {
 251         struct mm_struct *mm = current->mm;
 252         struct vm_area_struct *vma;
 253         struct hstate *h = hstate_file(file);
 254         const unsigned long mmap_end = arch_get_mmap_end(addr);
 255
 256         if (len & ~huge_page_mask(h))
 257                 return -EINVAL;
 258         if (len > TASK_SIZE)
 259                 return -ENOMEM;
 260
 261         if (flags & MAP_FIXED) {
 262                 if (prepare_hugepage_range(file, addr, len))
 263                         return -EINVAL;
 264                 return addr;
 265         }
 266
 267         if (addr) {
 268                 addr = ALIGN(addr, huge_page_size(h));
 269                 vma = find_vma(mm, addr);
 270                 if (mmap_end - len >= addr &&
 271                     (!vma || addr + len <= vm_start_gap(vma)))
 272                         return addr;
 273         }
 274
 275         /*
 276          * Use mm->get_unmapped_area value as a hint to use topdown routine.
 277          * If architectures have special needs, they should define their own
 278          * version of hugetlb_get_unmapped_area.
 279          */
 280         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
 281                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 282                                 pgoff, flags);
 283         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 284                         pgoff, flags);
 285 }
 286 #endif
 287
 288 static size_t
 289 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 290                         struct iov_iter *to, unsigned long size)
 291 {
 292         size_t copied = 0;
 293         int i, chunksize;
 294
 295         /* Find which 4k chunk and offset with in that chunk */
 296         i = offset >> PAGE_SHIFT;
 297         offset = offset & ~PAGE_MASK;
 298
 299         while (size) {
 300                 size_t n;
 301                 chunksize = PAGE_SIZE;
 302                 if (offset)
 303                         chunksize -= offset;
 304                 if (chunksize > size)
 305                         chunksize = size;
 306                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 307                 copied += n;
 308                 if (n != chunksize)
 309                         return copied;
 310                 offset = 0;
 311                 size -= chunksize;
 312                 i++;
 313         }
 314         return copied;
 315 }
 316
 317 /*
 318  * Support for read() - Find the page attached to f_mapping and copy out the
 319  * data. Its *very* similar to generic_file_buffered_read(), we can't use that
 320  * since it has PAGE_SIZE assumptions.
 321  */
 322 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 323 {
 324         struct file *file = iocb->ki_filp;
 325         struct hstate *h = hstate_file(file);
 326         struct address_space *mapping = file->f_mapping;
 327         struct inode *inode = mapping->host;
 328         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 329         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 330         unsigned long end_index;
 331         loff_t isize;
 332         ssize_t retval = 0;
 333
 334         while (iov_iter_count(to)) {
 335                 struct page *page;
 336                 size_t nr, copied;
 337
 338                 /* nr is the maximum number of bytes to copy from this page */
 339                 nr = huge_page_size(h);
 340                 isize = i_size_read(inode);
 341                 if (!isize)
 342                         break;
 343                 end_index = (isize - 1) >> huge_page_shift(h);
 344                 if (index > end_index)
 345                         break;
 346                 if (index == end_index) {
 347                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 348                         if (nr <= offset)
 349                                 break;
 350                 }
 351                 nr = nr - offset;
 352
 353                 /* Find the page */
 354                 page = find_lock_page(mapping, index);
 355                 if (unlikely(page == NULL)) {
 356                         /*
 357                          * We have a HOLE, zero out the user-buffer for the
 358                          * length of the hole or request.
 359                          */
 360                         copied = iov_iter_zero(nr, to);
 361                 } else {
 362                         unlock_page(page);
 363
 364                         if (PageHWPoison(page)) {
 365                                 put_page(page);
 366                                 retval = -EIO;
 367                                 break;
 368                         }
 369
 370                         /*
 371                          * We have the page, copy it to user space buffer.
 372                          */
 373                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 374                         put_page(page);
 375                 }
 376                 offset += copied;
 377                 retval += copied;
 378                 if (copied != nr && iov_iter_count(to)) {
 379                         if (!retval)
 380                                 retval = -EFAULT;
 381                         break;
 382                 }
 383                 index += offset >> huge_page_shift(h);
 384                 offset &= ~huge_page_mask(h);
 385         }
 386         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 387         return retval;
 388 }
 389
 390 static int hugetlbfs_write_begin(struct file *file,
 391                         struct address_space *mapping,
 392                         loff_t pos, unsigned len, unsigned flags,
 393                         struct page **pagep, void **fsdata)
 394 {
 395         return -EINVAL;
 396 }
 397
 398 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 399                         loff_t pos, unsigned len, unsigned copied,
 400                         struct page *page, void *fsdata)
 401 {
 402         BUG();
 403         return -EINVAL;
 404 }
 405
 406 static void remove_huge_page(struct page *page)
 407 {
 408         ClearPageDirty(page);
 409         ClearPageUptodate(page);
 410         delete_from_page_cache(page);
 411 }
 412
 413 static void
 414 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
 415 {
 416         struct vm_area_struct *vma;
 417
 418         /*
 419          * end == 0 indicates that the entire range after start should be
 420          * unmapped.  Note, end is exclusive, whereas the interval tree takes
 421          * an inclusive "last".
 422          */
 423         vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
 424                 unsigned long v_offset;
 425                 unsigned long v_end;
 426
 427                 /*
 428                  * Can the expression below overflow on 32-bit arches?
 429                  * No, because the interval tree returns us only those vmas
 430                  * which overlap the truncated area starting at pgoff,
 431                  * and no vma on a 32-bit arch can span beyond the 4GB.
 432                  */
 433                 if (vma->vm_pgoff < start)
 434                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 435                 else
 436                         v_offset = 0;
 437
 438                 if (!end)
 439                         v_end = vma->vm_end;
 440                 else {
 441                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 442                                                         + vma->vm_start;
 443                         if (v_end > vma->vm_end)
 444                                 v_end = vma->vm_end;
 445                 }
 446
 447                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 448                                                                         NULL);
 449         }
 450 }
 451
 452 /*
 453  * remove_inode_hugepages handles two distinct cases: truncation and hole
 454  * punch.  There are subtle differences in operation for each case.
 455  *
 456  * truncation is indicated by end of range being LLONG_MAX
 457  *      In this case, we first scan the range and release found pages.
 458  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
 459  *      maps and global counts.  Page faults can not race with truncation
 460  *      in this routine.  hugetlb_no_page() holds i_mmap_rwsem and prevents
 461  *      page faults in the truncated range by checking i_size.  i_size is
 462  *      modified while holding i_mmap_rwsem.
 463  * hole punch is indicated if end is not LLONG_MAX
 464  *      In the hole punch case we scan the range and release found pages.
 465  *      Only when releasing a page is the associated region/reserve map
 466  *      deleted.  The region/reserve map for ranges without associated
 467  *      pages are not modified.  Page faults can race with hole punch.
 468  *      This is indicated if we find a mapped page.
 469  * Note: If the passed end of range value is beyond the end of file, but
 470  * not LLONG_MAX this routine still performs a hole punch operation.
 471  */
 472 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 473                                    loff_t lend)
 474 {
 475         struct hstate *h = hstate_inode(inode);
 476         struct address_space *mapping = &inode->i_data;
 477         const pgoff_t start = lstart >> huge_page_shift(h);
 478         const pgoff_t end = lend >> huge_page_shift(h);
 479         struct pagevec pvec;
 480         pgoff_t next, index;
 481         int i, freed = 0;
 482         bool truncate_op = (lend == LLONG_MAX);
 483
 484         pagevec_init(&pvec);
 485         next = start;
 486         while (next < end) {
 487                 /*
 488                  * When no more pages are found, we are done.
 489                  */
 490                 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
 491                         break;
 492
 493                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 494                         struct page *page = pvec.pages[i];
 495                         u32 hash = 0;
 496
 497                         index = page->index;
 498                         if (!truncate_op) {
 499                                 /*
 500                                  * Only need to hold the fault mutex in the
 501                                  * hole punch case.  This prevents races with
 502                                  * page faults.  Races are not possible in the
 503                                  * case of truncation.
 504                                  */
 505                                 hash = hugetlb_fault_mutex_hash(mapping, index);
 506                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 507                         }
 508
 509                         /*
 510                          * If page is mapped, it was faulted in after being
 511                          * unmapped in caller.  Unmap (again) now after taking
 512                          * the fault mutex.  The mutex will prevent faults
 513                          * until we finish removing the page.
 514                          *
 515                          * This race can only happen in the hole punch case.
 516                          * Getting here in a truncate operation is a bug.
 517                          */
 518                         if (unlikely(page_mapped(page))) {
 519                                 BUG_ON(truncate_op);
 520
 521                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 522                                 i_mmap_lock_write(mapping);
 523                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 524                                 hugetlb_vmdelete_list(&mapping->i_mmap,
 525                                         index * pages_per_huge_page(h),
 526                                         (index + 1) * pages_per_huge_page(h));
 527                                 i_mmap_unlock_write(mapping);
 528                         }
 529
 530                         lock_page(page);
 531                         /*
 532                          * We must free the huge page and remove from page
 533                          * cache (remove_huge_page) BEFORE removing the
 534                          * region/reserve map (hugetlb_unreserve_pages).  In
 535                          * rare out of memory conditions, removal of the
 536                          * region/reserve map could fail. Correspondingly,
 537                          * the subpool and global reserve usage count can need
 538                          * to be adjusted.
 539                          */
 540                         VM_BUG_ON(HPageRestoreReserve(page));
 541                         remove_huge_page(page);
 542                         freed++;
 543                         if (!truncate_op) {
 544                                 if (unlikely(hugetlb_unreserve_pages(inode,
 545                                                         index, index + 1, 1)))
 546                                         hugetlb_fix_reserve_counts(inode);
 547                         }
 548
 549                         unlock_page(page);
 550                         if (!truncate_op)
 551                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 552                 }
 553                 huge_pagevec_release(&pvec);
 554                 cond_resched();
 555         }
 556
 557         if (truncate_op)
 558                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 559 }
 560
 561 static void hugetlbfs_evict_inode(struct inode *inode)
 562 {
 563         struct resv_map *resv_map;
 564
 565         remove_inode_hugepages(inode, 0, LLONG_MAX);
 566
 567         /*
 568          * Get the resv_map from the address space embedded in the inode.
 569          * This is the address space which points to any resv_map allocated
 570          * at inode creation time.  If this is a device special inode,
 571          * i_mapping may not point to the original address space.
 572          */
 573         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
 574         /* Only regular and link inodes have associated reserve maps */
 575         if (resv_map)
 576                 resv_map_release(&resv_map->refs);
 577         clear_inode(inode);
 578 }
 579
 580 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 581 {
 582         pgoff_t pgoff;
 583         struct address_space *mapping = inode->i_mapping;
 584         struct hstate *h = hstate_inode(inode);
 585
 586         BUG_ON(offset & ~huge_page_mask(h));
 587         pgoff = offset >> PAGE_SHIFT;
 588
 589         i_mmap_lock_write(mapping);
 590         i_size_write(inode, offset);
 591         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 592                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 593         i_mmap_unlock_write(mapping);
 594         remove_inode_hugepages(inode, offset, LLONG_MAX);
 595 }
 596
 597 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 598 {
 599         struct hstate *h = hstate_inode(inode);
 600         loff_t hpage_size = huge_page_size(h);
 601         loff_t hole_start, hole_end;
 602
 603         /*
 604          * For hole punch round up the beginning offset of the hole and
 605          * round down the end.
 606          */
 607         hole_start = round_up(offset, hpage_size);
 608         hole_end = round_down(offset + len, hpage_size);
 609
 610         if (hole_end > hole_start) {
 611                 struct address_space *mapping = inode->i_mapping;
 612                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 613
 614                 inode_lock(inode);
 615
 616                 /* protected by i_rwsem */
 617                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 618                         inode_unlock(inode);
 619                         return -EPERM;
 620                 }
 621
 622                 i_mmap_lock_write(mapping);
 623                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 624                         hugetlb_vmdelete_list(&mapping->i_mmap,
 625                                                 hole_start >> PAGE_SHIFT,
 626                                                 hole_end  >> PAGE_SHIFT);
 627                 i_mmap_unlock_write(mapping);
 628                 remove_inode_hugepages(inode, hole_start, hole_end);
 629                 inode_unlock(inode);
 630         }
 631
 632         return 0;
 633 }
 634
 635 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 636                                 loff_t len)
 637 {
 638         struct inode *inode = file_inode(file);
 639         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 640         struct address_space *mapping = inode->i_mapping;
 641         struct hstate *h = hstate_inode(inode);
 642         struct vm_area_struct pseudo_vma;
 643         struct mm_struct *mm = current->mm;
 644         loff_t hpage_size = huge_page_size(h);
 645         unsigned long hpage_shift = huge_page_shift(h);
 646         pgoff_t start, index, end;
 647         int error;
 648         u32 hash;
 649
 650         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 651                 return -EOPNOTSUPP;
 652
 653         if (mode & FALLOC_FL_PUNCH_HOLE)
 654                 return hugetlbfs_punch_hole(inode, offset, len);
 655
 656         /*
 657          * Default preallocate case.
 658          * For this range, start is rounded down and end is rounded up
 659          * as well as being converted to page offsets.
 660          */
 661         start = offset >> hpage_shift;
 662         end = (offset + len + hpage_size - 1) >> hpage_shift;
 663
 664         inode_lock(inode);
 665
 666         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 667         error = inode_newsize_ok(inode, offset + len);
 668         if (error)
 669                 goto out;
 670
 671         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 672                 error = -EPERM;
 673                 goto out;
 674         }
 675
 676         /*
 677          * Initialize a pseudo vma as this is required by the huge page
 678          * allocation routines.  If NUMA is configured, use page index
 679          * as input to create an allocation policy.
 680          */
 681         vma_init(&pseudo_vma, mm);
 682         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 683         pseudo_vma.vm_file = file;
 684
 685         for (index = start; index < end; index++) {
 686                 /*
 687                  * This is supposed to be the vaddr where the page is being
 688                  * faulted in, but we have no vaddr here.
 689                  */
 690                 struct page *page;
 691                 unsigned long addr;
 692
 693                 cond_resched();
 694
 695                 /*
 696                  * fallocate(2) manpage permits EINTR; we may have been
 697                  * interrupted because we are using up too much memory.
 698                  */
 699                 if (signal_pending(current)) {
 700                         error = -EINTR;
 701                         break;
 702                 }
 703
 704                 /* Set numa allocation policy based on index */
 705                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 706
 707                 /* addr is the offset within the file (zero based) */
 708                 addr = index * hpage_size;
 709
 710                 /*
 711                  * fault mutex taken here, protects against fault path
 712                  * and hole punch.  inode_lock previously taken protects
 713                  * against truncation.
 714                  */
 715                 hash = hugetlb_fault_mutex_hash(mapping, index);
 716                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 717
 718                 /* See if already present in mapping to avoid alloc/free */
 719                 page = find_get_page(mapping, index);
 720                 if (page) {
 721                         put_page(page);
 722                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 723                         hugetlb_drop_vma_policy(&pseudo_vma);
 724                         continue;
 725                 }
 726
 727                 /*
 728                  * Allocate page without setting the avoid_reserve argument.
 729                  * There certainly are no reserves associated with the
 730                  * pseudo_vma.  However, there could be shared mappings with
 731                  * reserves for the file at the inode level.  If we fallocate
 732                  * pages in these areas, we need to consume the reserves
 733                  * to keep reservation accounting consistent.
 734                  */
 735                 page = alloc_huge_page(&pseudo_vma, addr, 0);
 736                 hugetlb_drop_vma_policy(&pseudo_vma);
 737                 if (IS_ERR(page)) {
 738                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 739                         error = PTR_ERR(page);
 740                         goto out;
 741                 }
 742                 clear_huge_page(page, addr, pages_per_huge_page(h));
 743                 __SetPageUptodate(page);
 744                 error = huge_add_to_page_cache(page, mapping, index);
 745                 if (unlikely(error)) {
 746                         restore_reserve_on_error(h, &pseudo_vma, addr, page);
 747                         put_page(page);
 748                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 749                         goto out;
 750                 }
 751
 752                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 753
 754                 SetHPageMigratable(page);
 755                 /*
 756                  * unlock_page because locked by add_to_page_cache()
 757                  * put_page() due to reference from alloc_huge_page()
 758                  */
 759                 unlock_page(page);
 760                 put_page(page);
 761         }
 762
 763         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 764                 i_size_write(inode, offset + len);
 765         inode->i_ctime = current_time(inode);
 766 out:
 767         inode_unlock(inode);
 768         return error;
 769 }
 770
 771 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
 772                              struct dentry *dentry, struct iattr *attr)
 773 {
 774         struct inode *inode = d_inode(dentry);
 775         struct hstate *h = hstate_inode(inode);
 776         int error;
 777         unsigned int ia_valid = attr->ia_valid;
 778         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 779
 780         error = setattr_prepare(&init_user_ns, dentry, attr);
 781         if (error)
 782                 return error;
 783
 784         if (ia_valid & ATTR_SIZE) {
 785                 loff_t oldsize = inode->i_size;
 786                 loff_t newsize = attr->ia_size;
 787
 788                 if (newsize & ~huge_page_mask(h))
 789                         return -EINVAL;
 790                 /* protected by i_rwsem */
 791                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 792                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 793                         return -EPERM;
 794                 hugetlb_vmtruncate(inode, newsize);
 795         }
 796
 797         setattr_copy(&init_user_ns, inode, attr);
 798         mark_inode_dirty(inode);
 799         return 0;
 800 }
 801
 802 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 803                                         struct hugetlbfs_fs_context *ctx)
 804 {
 805         struct inode *inode;
 806
 807         inode = new_inode(sb);
 808         if (inode) {
 809                 inode->i_ino = get_next_ino();
 810                 inode->i_mode = S_IFDIR | ctx->mode;
 811                 inode->i_uid = ctx->uid;
 812                 inode->i_gid = ctx->gid;
 813                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 814                 inode->i_op = &hugetlbfs_dir_inode_operations;
 815                 inode->i_fop = &simple_dir_operations;
 816                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 817                 inc_nlink(inode);
 818                 lockdep_annotate_inode_mutex_key(inode);
 819         }
 820         return inode;
 821 }
 822
 823 /*
 824  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 825  * be taken from reclaim -- unlike regular filesystems. This needs an
 826  * annotation because huge_pmd_share() does an allocation under hugetlb's
 827  * i_mmap_rwsem.
 828  */
 829 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 830
 831 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 832                                         struct inode *dir,
 833                                         umode_t mode, dev_t dev)
 834 {
 835         struct inode *inode;
 836         struct resv_map *resv_map = NULL;
 837
 838         /*
 839          * Reserve maps are only needed for inodes that can have associated
 840          * page allocations.
 841          */
 842         if (S_ISREG(mode) || S_ISLNK(mode)) {
 843                 resv_map = resv_map_alloc();
 844                 if (!resv_map)
 845                         return NULL;
 846         }
 847
 848         inode = new_inode(sb);
 849         if (inode) {
 850                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 851
 852                 inode->i_ino = get_next_ino();
 853                 inode_init_owner(&init_user_ns, inode, dir, mode);
 854                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 855                                 &hugetlbfs_i_mmap_rwsem_key);
 856                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 857                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 858                 inode->i_mapping->private_data = resv_map;
 859                 info->seals = F_SEAL_SEAL;
 860                 switch (mode & S_IFMT) {
 861                 default:
 862                         init_special_inode(inode, mode, dev);
 863                         break;
 864                 case S_IFREG:
 865                         inode->i_op = &hugetlbfs_inode_operations;
 866                         inode->i_fop = &hugetlbfs_file_operations;
 867                         break;
 868                 case S_IFDIR:
 869                         inode->i_op = &hugetlbfs_dir_inode_operations;
 870                         inode->i_fop = &simple_dir_operations;
 871
 872                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 873                         inc_nlink(inode);
 874                         break;
 875                 case S_IFLNK:
 876                         inode->i_op = &page_symlink_inode_operations;
 877                         inode_nohighmem(inode);
 878                         break;
 879                 }
 880                 lockdep_annotate_inode_mutex_key(inode);
 881         } else {
 882                 if (resv_map)
 883                         kref_put(&resv_map->refs, resv_map_release);
 884         }
 885
 886         return inode;
 887 }
 888
 889 /*
 890  * File creation. Allocate an inode, and we're done..
 891  */
 892 static int do_hugetlbfs_mknod(struct inode *dir,
 893                         struct dentry *dentry,
 894                         umode_t mode,
 895                         dev_t dev,
 896                         bool tmpfile)
 897 {
 898         struct inode *inode;
 899         int error = -ENOSPC;
 900
 901         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 902         if (inode) {
 903                 dir->i_ctime = dir->i_mtime = current_time(dir);
 904                 if (tmpfile) {
 905                         d_tmpfile(dentry, inode);
 906                 } else {
 907                         d_instantiate(dentry, inode);
 908                         dget(dentry);/* Extra count - pin the dentry in core */
 909                 }
 910                 error = 0;
 911         }
 912         return error;
 913 }
 914
 915 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
 916                            struct dentry *dentry, umode_t mode, dev_t dev)
 917 {
 918         return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
 919 }
 920
 921 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
 922                            struct dentry *dentry, umode_t mode)
 923 {
 924         int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
 925                                      mode | S_IFDIR, 0);
 926         if (!retval)
 927                 inc_nlink(dir);
 928         return retval;
 929 }
 930
 931 static int hugetlbfs_create(struct user_namespace *mnt_userns,
 932                             struct inode *dir, struct dentry *dentry,
 933                             umode_t mode, bool excl)
 934 {
 935         return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
 936 }
 937
 938 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
 939                              struct inode *dir, struct dentry *dentry,
 940                              umode_t mode)
 941 {
 942         return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
 943 }
 944
 945 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
 946                              struct inode *dir, struct dentry *dentry,
 947                              const char *symname)
 948 {
 949         struct inode *inode;
 950         int error = -ENOSPC;
 951
 952         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 953         if (inode) {
 954                 int l = strlen(symname)+1;
 955                 error = page_symlink(inode, symname, l);
 956                 if (!error) {
 957                         d_instantiate(dentry, inode);
 958                         dget(dentry);
 959                 } else
 960                         iput(inode);
 961         }
 962         dir->i_ctime = dir->i_mtime = current_time(dir);
 963
 964         return error;
 965 }
 966
 967 static int hugetlbfs_migrate_page(struct address_space *mapping,
 968                                 struct page *newpage, struct page *page,
 969                                 enum migrate_mode mode)
 970 {
 971         int rc;
 972
 973         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 974         if (rc != MIGRATEPAGE_SUCCESS)
 975                 return rc;
 976
 977         if (hugetlb_page_subpool(page)) {
 978                 hugetlb_set_page_subpool(newpage, hugetlb_page_subpool(page));
 979                 hugetlb_set_page_subpool(page, NULL);
 980         }
 981
 982         if (mode != MIGRATE_SYNC_NO_COPY)
 983                 migrate_page_copy(newpage, page);
 984         else
 985                 migrate_page_states(newpage, page);
 986
 987         return MIGRATEPAGE_SUCCESS;
 988 }
 989
 990 static int hugetlbfs_error_remove_page(struct address_space *mapping,
 991                                 struct page *page)
 992 {
 993         return 0;
 994 }
 995
 996 /*
 997  * Display the mount options in /proc/mounts.
 998  */
 999 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1000 {
1001         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1002         struct hugepage_subpool *spool = sbinfo->spool;
1003         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1004         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1005         char mod;
1006
1007         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1008                 seq_printf(m, ",uid=%u",
1009                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1010         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1011                 seq_printf(m, ",gid=%u",
1012                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1013         if (sbinfo->mode != 0755)
1014                 seq_printf(m, ",mode=%o", sbinfo->mode);
1015         if (sbinfo->max_inodes != -1)
1016                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1017
1018         hpage_size /= 1024;
1019         mod = 'K';
1020         if (hpage_size >= 1024) {
1021                 hpage_size /= 1024;
1022                 mod = 'M';
1023         }
1024         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1025         if (spool) {
1026                 if (spool->max_hpages != -1)
1027                         seq_printf(m, ",size=%llu",
1028                                    (unsigned long long)spool->max_hpages << hpage_shift);
1029                 if (spool->min_hpages != -1)
1030                         seq_printf(m, ",min_size=%llu",
1031                                    (unsigned long long)spool->min_hpages << hpage_shift);
1032         }
1033         return 0;
1034 }
1035
1036 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1037 {
1038         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1039         struct hstate *h = hstate_inode(d_inode(dentry));
1040
1041         buf->f_type = HUGETLBFS_MAGIC;
1042         buf->f_bsize = huge_page_size(h);
1043         if (sbinfo) {
1044                 spin_lock(&sbinfo->stat_lock);
1045                 /* If no limits set, just report 0 for max/free/used
1046                  * blocks, like simple_statfs() */
1047                 if (sbinfo->spool) {
1048                         long free_pages;
1049
1050                         spin_lock_irq(&sbinfo->spool->lock);
1051                         buf->f_blocks = sbinfo->spool->max_hpages;
1052                         free_pages = sbinfo->spool->max_hpages
1053                                 - sbinfo->spool->used_hpages;
1054                         buf->f_bavail = buf->f_bfree = free_pages;
1055                         spin_unlock_irq(&sbinfo->spool->lock);
1056                         buf->f_files = sbinfo->max_inodes;
1057                         buf->f_ffree = sbinfo->free_inodes;
1058                 }
1059                 spin_unlock(&sbinfo->stat_lock);
1060         }
1061         buf->f_namelen = NAME_MAX;
1062         return 0;
1063 }
1064
1065 static void hugetlbfs_put_super(struct super_block *sb)
1066 {
1067         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1068
1069         if (sbi) {
1070                 sb->s_fs_info = NULL;
1071
1072                 if (sbi->spool)
1073                         hugepage_put_subpool(sbi->spool);
1074
1075                 kfree(sbi);
1076         }
1077 }
1078
1079 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1080 {
1081         if (sbinfo->free_inodes >= 0) {
1082                 spin_lock(&sbinfo->stat_lock);
1083                 if (unlikely(!sbinfo->free_inodes)) {
1084                         spin_unlock(&sbinfo->stat_lock);
1085                         return 0;
1086                 }
1087                 sbinfo->free_inodes--;
1088                 spin_unlock(&sbinfo->stat_lock);
1089         }
1090
1091         return 1;
1092 }
1093
1094 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1095 {
1096         if (sbinfo->free_inodes >= 0) {
1097                 spin_lock(&sbinfo->stat_lock);
1098                 sbinfo->free_inodes++;
1099                 spin_unlock(&sbinfo->stat_lock);
1100         }
1101 }
1102
1103
1104 static struct kmem_cache *hugetlbfs_inode_cachep;
1105
1106 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1107 {
1108         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1109         struct hugetlbfs_inode_info *p;
1110
1111         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1112                 return NULL;
1113         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1114         if (unlikely(!p)) {
1115                 hugetlbfs_inc_free_inodes(sbinfo);
1116                 return NULL;
1117         }
1118
1119         /*
1120          * Any time after allocation, hugetlbfs_destroy_inode can be called
1121          * for the inode.  mpol_free_shared_policy is unconditionally called
1122          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1123          * in case of a quick call to destroy.
1124          *
1125          * Note that the policy is initialized even if we are creating a
1126          * private inode.  This simplifies hugetlbfs_destroy_inode.
1127          */
1128         mpol_shared_policy_init(&p->policy, NULL);
1129
1130         return &p->vfs_inode;
1131 }
1132
1133 static void hugetlbfs_free_inode(struct inode *inode)
1134 {
1135         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1136 }
1137
1138 static void hugetlbfs_destroy_inode(struct inode *inode)
1139 {
1140         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1141         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1142 }
1143
1144 static const struct address_space_operations hugetlbfs_aops = {
1145         .write_begin    = hugetlbfs_write_begin,
1146         .write_end      = hugetlbfs_write_end,
1147         .set_page_dirty =  __set_page_dirty_no_writeback,
1148         .migratepage    = hugetlbfs_migrate_page,
1149         .error_remove_page      = hugetlbfs_error_remove_page,
1150 };
1151
1152
1153 static void init_once(void *foo)
1154 {
1155         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1156
1157         inode_init_once(&ei->vfs_inode);
1158 }
1159
1160 const struct file_operations hugetlbfs_file_operations = {
1161         .read_iter              = hugetlbfs_read_iter,
1162         .mmap                   = hugetlbfs_file_mmap,
1163         .fsync                  = noop_fsync,
1164         .get_unmapped_area      = hugetlb_get_unmapped_area,
1165         .llseek                 = default_llseek,
1166         .fallocate              = hugetlbfs_fallocate,
1167 };
1168
1169 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1170         .create         = hugetlbfs_create,
1171         .lookup         = simple_lookup,
1172         .link           = simple_link,
1173         .unlink         = simple_unlink,
1174         .symlink        = hugetlbfs_symlink,
1175         .mkdir          = hugetlbfs_mkdir,
1176         .rmdir          = simple_rmdir,
1177         .mknod          = hugetlbfs_mknod,
1178         .rename         = simple_rename,
1179         .setattr        = hugetlbfs_setattr,
1180         .tmpfile        = hugetlbfs_tmpfile,
1181 };
1182
1183 static const struct inode_operations hugetlbfs_inode_operations = {
1184         .setattr        = hugetlbfs_setattr,
1185 };
1186
1187 static const struct super_operations hugetlbfs_ops = {
1188         .alloc_inode    = hugetlbfs_alloc_inode,
1189         .free_inode     = hugetlbfs_free_inode,
1190         .destroy_inode  = hugetlbfs_destroy_inode,
1191         .evict_inode    = hugetlbfs_evict_inode,
1192         .statfs         = hugetlbfs_statfs,
1193         .put_super      = hugetlbfs_put_super,
1194         .show_options   = hugetlbfs_show_options,
1195 };
1196
1197 /*
1198  * Convert size option passed from command line to number of huge pages
1199  * in the pool specified by hstate.  Size option could be in bytes
1200  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1201  */
1202 static long
1203 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1204                          enum hugetlbfs_size_type val_type)
1205 {
1206         if (val_type == NO_SIZE)
1207                 return -1;
1208
1209         if (val_type == SIZE_PERCENT) {
1210                 size_opt <<= huge_page_shift(h);
1211                 size_opt *= h->max_huge_pages;
1212                 do_div(size_opt, 100);
1213         }
1214
1215         size_opt >>= huge_page_shift(h);
1216         return size_opt;
1217 }
1218
1219 /*
1220  * Parse one mount parameter.
1221  */
1222 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1223 {
1224         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1225         struct fs_parse_result result;
1226         char *rest;
1227         unsigned long ps;
1228         int opt;
1229
1230         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1231         if (opt < 0)
1232                 return opt;
1233
1234         switch (opt) {
1235         case Opt_uid:
1236                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1237                 if (!uid_valid(ctx->uid))
1238                         goto bad_val;
1239                 return 0;
1240
1241         case Opt_gid:
1242                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1243                 if (!gid_valid(ctx->gid))
1244                         goto bad_val;
1245                 return 0;
1246
1247         case Opt_mode:
1248                 ctx->mode = result.uint_32 & 01777U;
1249                 return 0;
1250
1251         case Opt_size:
1252                 /* memparse() will accept a K/M/G without a digit */
1253                 if (!param->string || !isdigit(param->string[0]))
1254                         goto bad_val;
1255                 ctx->max_size_opt = memparse(param->string, &rest);
1256                 ctx->max_val_type = SIZE_STD;
1257                 if (*rest == '%')
1258                         ctx->max_val_type = SIZE_PERCENT;
1259                 return 0;
1260
1261         case Opt_nr_inodes:
1262                 /* memparse() will accept a K/M/G without a digit */
1263                 if (!param->string || !isdigit(param->string[0]))
1264                         goto bad_val;
1265                 ctx->nr_inodes = memparse(param->string, &rest);
1266                 return 0;
1267
1268         case Opt_pagesize:
1269                 ps = memparse(param->string, &rest);
1270                 ctx->hstate = size_to_hstate(ps);
1271                 if (!ctx->hstate) {
1272                         pr_err("Unsupported page size %lu MB\n", ps >> 20);
1273                         return -EINVAL;
1274                 }
1275                 return 0;
1276
1277         case Opt_min_size:
1278                 /* memparse() will accept a K/M/G without a digit */
1279                 if (!param->string || !isdigit(param->string[0]))
1280                         goto bad_val;
1281                 ctx->min_size_opt = memparse(param->string, &rest);
1282                 ctx->min_val_type = SIZE_STD;
1283                 if (*rest == '%')
1284                         ctx->min_val_type = SIZE_PERCENT;
1285                 return 0;
1286
1287         default:
1288                 return -EINVAL;
1289         }
1290
1291 bad_val:
1292         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1293                       param->string, param->key);
1294 }
1295
1296 /*
1297  * Validate the parsed options.
1298  */
1299 static int hugetlbfs_validate(struct fs_context *fc)
1300 {
1301         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1302
1303         /*
1304          * Use huge page pool size (in hstate) to convert the size
1305          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1306          */
1307         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1308                                                    ctx->max_size_opt,
1309                                                    ctx->max_val_type);
1310         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1311                                                    ctx->min_size_opt,
1312                                                    ctx->min_val_type);
1313
1314         /*
1315          * If max_size was specified, then min_size must be smaller
1316          */
1317         if (ctx->max_val_type > NO_SIZE &&
1318             ctx->min_hpages > ctx->max_hpages) {
1319                 pr_err("Minimum size can not be greater than maximum size\n");
1320                 return -EINVAL;
1321         }
1322
1323         return 0;
1324 }
1325
1326 static int
1327 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1328 {
1329         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1330         struct hugetlbfs_sb_info *sbinfo;
1331
1332         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1333         if (!sbinfo)
1334                 return -ENOMEM;
1335         sb->s_fs_info = sbinfo;
1336         spin_lock_init(&sbinfo->stat_lock);
1337         sbinfo->hstate          = ctx->hstate;
1338         sbinfo->max_inodes      = ctx->nr_inodes;
1339         sbinfo->free_inodes     = ctx->nr_inodes;
1340         sbinfo->spool           = NULL;
1341         sbinfo->uid             = ctx->uid;
1342         sbinfo->gid             = ctx->gid;
1343         sbinfo->mode            = ctx->mode;
1344
1345         /*
1346          * Allocate and initialize subpool if maximum or minimum size is
1347          * specified.  Any needed reservations (for minimum size) are taken
1348          * taken when the subpool is created.
1349          */
1350         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1351                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1352                                                      ctx->max_hpages,
1353                                                      ctx->min_hpages);
1354                 if (!sbinfo->spool)
1355                         goto out_free;
1356         }
1357         sb->s_maxbytes = MAX_LFS_FILESIZE;
1358         sb->s_blocksize = huge_page_size(ctx->hstate);
1359         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1360         sb->s_magic = HUGETLBFS_MAGIC;
1361         sb->s_op = &hugetlbfs_ops;
1362         sb->s_time_gran = 1;
1363
1364         /*
1365          * Due to the special and limited functionality of hugetlbfs, it does
1366          * not work well as a stacking filesystem.
1367          */
1368         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1369         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1370         if (!sb->s_root)
1371                 goto out_free;
1372         return 0;
1373 out_free:
1374         kfree(sbinfo->spool);
1375         kfree(sbinfo);
1376         return -ENOMEM;
1377 }
1378
1379 static int hugetlbfs_get_tree(struct fs_context *fc)
1380 {
1381         int err = hugetlbfs_validate(fc);
1382         if (err)
1383                 return err;
1384         return get_tree_nodev(fc, hugetlbfs_fill_super);
1385 }
1386
1387 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1388 {
1389         kfree(fc->fs_private);
1390 }
1391
1392 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1393         .free           = hugetlbfs_fs_context_free,
1394         .parse_param    = hugetlbfs_parse_param,
1395         .get_tree       = hugetlbfs_get_tree,
1396 };
1397
1398 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1399 {
1400         struct hugetlbfs_fs_context *ctx;
1401
1402         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1403         if (!ctx)
1404                 return -ENOMEM;
1405
1406         ctx->max_hpages = -1; /* No limit on size by default */
1407         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1408         ctx->uid        = current_fsuid();
1409         ctx->gid        = current_fsgid();
1410         ctx->mode       = 0755;
1411         ctx->hstate     = &default_hstate;
1412         ctx->min_hpages = -1; /* No default minimum size */
1413         ctx->max_val_type = NO_SIZE;
1414         ctx->min_val_type = NO_SIZE;
1415         fc->fs_private = ctx;
1416         fc->ops = &hugetlbfs_fs_context_ops;
1417         return 0;
1418 }
1419
1420 static struct file_system_type hugetlbfs_fs_type = {
1421         .name                   = "hugetlbfs",
1422         .init_fs_context        = hugetlbfs_init_fs_context,
1423         .parameters             = hugetlb_fs_parameters,
1424         .kill_sb                = kill_litter_super,
1425 };
1426
1427 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1428
1429 static int can_do_hugetlb_shm(void)
1430 {
1431         kgid_t shm_group;
1432         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1433         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1434 }
1435
1436 static int get_hstate_idx(int page_size_log)
1437 {
1438         struct hstate *h = hstate_sizelog(page_size_log);
1439
1440         if (!h)
1441                 return -1;
1442         return hstate_index(h);
1443 }
1444
1445 /*
1446  * Note that size should be aligned to proper hugepage size in caller side,
1447  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1448  */
1449 struct file *hugetlb_file_setup(const char *name, size_t size,
1450                                 vm_flags_t acctflag, struct ucounts **ucounts,
1451                                 int creat_flags, int page_size_log)
1452 {
1453         struct inode *inode;
1454         struct vfsmount *mnt;
1455         int hstate_idx;
1456         struct file *file;
1457
1458         hstate_idx = get_hstate_idx(page_size_log);
1459         if (hstate_idx < 0)
1460                 return ERR_PTR(-ENODEV);
1461
1462         *ucounts = NULL;
1463         mnt = hugetlbfs_vfsmount[hstate_idx];
1464         if (!mnt)
1465                 return ERR_PTR(-ENOENT);
1466
1467         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1468                 *ucounts = current_ucounts();
1469                 if (user_shm_lock(size, *ucounts)) {
1470                         task_lock(current);
1471                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1472                                 current->comm, current->pid);
1473                         task_unlock(current);
1474                 } else {
1475                         *ucounts = NULL;
1476                         return ERR_PTR(-EPERM);
1477                 }
1478         }
1479
1480         file = ERR_PTR(-ENOSPC);
1481         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1482         if (!inode)
1483                 goto out;
1484         if (creat_flags == HUGETLB_SHMFS_INODE)
1485                 inode->i_flags |= S_PRIVATE;
1486
1487         inode->i_size = size;
1488         clear_nlink(inode);
1489
1490         if (!hugetlb_reserve_pages(inode, 0,
1491                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1492                         acctflag))
1493                 file = ERR_PTR(-ENOMEM);
1494         else
1495                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1496                                         &hugetlbfs_file_operations);
1497         if (!IS_ERR(file))
1498                 return file;
1499
1500         iput(inode);
1501 out:
1502         if (*ucounts) {
1503                 user_shm_unlock(size, *ucounts);
1504                 *ucounts = NULL;
1505         }
1506         return file;
1507 }
1508
1509 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1510 {
1511         struct fs_context *fc;
1512         struct vfsmount *mnt;
1513
1514         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1515         if (IS_ERR(fc)) {
1516                 mnt = ERR_CAST(fc);
1517         } else {
1518                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1519                 ctx->hstate = h;
1520                 mnt = fc_mount(fc);
1521                 put_fs_context(fc);
1522         }
1523         if (IS_ERR(mnt))
1524                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1525                        huge_page_size(h) >> 10);
1526         return mnt;
1527 }
1528
1529 static int __init init_hugetlbfs_fs(void)
1530 {
1531         struct vfsmount *mnt;
1532         struct hstate *h;
1533         int error;
1534         int i;
1535
1536         if (!hugepages_supported()) {
1537                 pr_info("disabling because there are no supported hugepage sizes\n");
1538                 return -ENOTSUPP;
1539         }
1540
1541         error = -ENOMEM;
1542         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1543                                         sizeof(struct hugetlbfs_inode_info),
1544                                         0, SLAB_ACCOUNT, init_once);
1545         if (hugetlbfs_inode_cachep == NULL)
1546                 goto out;
1547
1548         error = register_filesystem(&hugetlbfs_fs_type);
1549         if (error)
1550                 goto out_free;
1551
1552         /* default hstate mount is required */
1553         mnt = mount_one_hugetlbfs(&default_hstate);
1554         if (IS_ERR(mnt)) {
1555                 error = PTR_ERR(mnt);
1556                 goto out_unreg;
1557         }
1558         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1559
1560         /* other hstates are optional */
1561         i = 0;
1562         for_each_hstate(h) {
1563                 if (i == default_hstate_idx) {
1564                         i++;
1565                         continue;
1566                 }
1567
1568                 mnt = mount_one_hugetlbfs(h);
1569                 if (IS_ERR(mnt))
1570                         hugetlbfs_vfsmount[i] = NULL;
1571                 else
1572                         hugetlbfs_vfsmount[i] = mnt;
1573                 i++;
1574         }
1575
1576         return 0;
1577
1578  out_unreg:
1579         (void)unregister_filesystem(&hugetlbfs_fs_type);
1580  out_free:
1581         kmem_cache_destroy(hugetlbfs_inode_cachep);
1582  out:
1583         return error;
1584 }
1585 fs_initcall(init_hugetlbfs_fs)