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 /*
 112  * Mask used when checking the page offset value passed in via system
 113  * calls.  This value will be converted to a loff_t which is signed.
 114  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 115  * value.  The extra bit (- 1 in the shift value) is to take the sign
 116  * bit into account.
 117  */
 118 #define PGOFF_LOFFT_MAX \
 119         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 120
 121 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 122 {
 123         struct inode *inode = file_inode(file);
 124         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 125         loff_t len, vma_len;
 126         int ret;
 127         struct hstate *h = hstate_file(file);
 128
 129         /*
 130          * vma address alignment (but not the pgoff alignment) has
 131          * already been checked by prepare_hugepage_range.  If you add
 132          * any error returns here, do so after setting VM_HUGETLB, so
 133          * is_vm_hugetlb_page tests below unmap_region go the right
 134          * way when do_mmap unwinds (may be important on powerpc
 135          * and ia64).
 136          */
 137         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 138         vma->vm_ops = &hugetlb_vm_ops;
 139
 140         ret = seal_check_future_write(info->seals, vma);
 141         if (ret)
 142                 return ret;
 143
 144         /*
 145          * page based offset in vm_pgoff could be sufficiently large to
 146          * overflow a loff_t when converted to byte offset.  This can
 147          * only happen on architectures where sizeof(loff_t) ==
 148          * sizeof(unsigned long).  So, only check in those instances.
 149          */
 150         if (sizeof(unsigned long) == sizeof(loff_t)) {
 151                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 152                         return -EINVAL;
 153         }
 154
 155         /* must be huge page aligned */
 156         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 157                 return -EINVAL;
 158
 159         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 160         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 161         /* check for overflow */
 162         if (len < vma_len)
 163                 return -EINVAL;
 164
 165         inode_lock(inode);
 166         file_accessed(file);
 167
 168         ret = -ENOMEM;
 169         if (!hugetlb_reserve_pages(inode,
 170                                 vma->vm_pgoff >> huge_page_order(h),
 171                                 len >> huge_page_shift(h), vma,
 172                                 vma->vm_flags))
 173                 goto out;
 174
 175         ret = 0;
 176         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 177                 i_size_write(inode, len);
 178 out:
 179         inode_unlock(inode);
 180
 181         return ret;
 182 }
 183
 184 /*
 185  * Called under mmap_write_lock(mm).
 186  */
 187
 188 static unsigned long
 189 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
 190                 unsigned long len, unsigned long pgoff, unsigned long flags)
 191 {
 192         struct hstate *h = hstate_file(file);
 193         struct vm_unmapped_area_info info;
 194
 195         info.flags = 0;
 196         info.length = len;
 197         info.low_limit = current->mm->mmap_base;
 198         info.high_limit = arch_get_mmap_end(addr, len, flags);
 199         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 200         info.align_offset = 0;
 201         return vm_unmapped_area(&info);
 202 }
 203
 204 static unsigned long
 205 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
 206                 unsigned long len, unsigned long pgoff, unsigned long flags)
 207 {
 208         struct hstate *h = hstate_file(file);
 209         struct vm_unmapped_area_info info;
 210
 211         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 212         info.length = len;
 213         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
 214         info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
 215         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 216         info.align_offset = 0;
 217         addr = vm_unmapped_area(&info);
 218
 219         /*
 220          * A failed mmap() very likely causes application failure,
 221          * so fall back to the bottom-up function here. This scenario
 222          * can happen with large stack limits and large mmap()
 223          * allocations.
 224          */
 225         if (unlikely(offset_in_page(addr))) {
 226                 VM_BUG_ON(addr != -ENOMEM);
 227                 info.flags = 0;
 228                 info.low_limit = current->mm->mmap_base;
 229                 info.high_limit = arch_get_mmap_end(addr, len, flags);
 230                 addr = vm_unmapped_area(&info);
 231         }
 232
 233         return addr;
 234 }
 235
 236 unsigned long
 237 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 238                                   unsigned long len, unsigned long pgoff,
 239                                   unsigned long flags)
 240 {
 241         struct mm_struct *mm = current->mm;
 242         struct vm_area_struct *vma;
 243         struct hstate *h = hstate_file(file);
 244         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
 245
 246         if (len & ~huge_page_mask(h))
 247                 return -EINVAL;
 248         if (len > TASK_SIZE)
 249                 return -ENOMEM;
 250
 251         if (flags & MAP_FIXED) {
 252                 if (prepare_hugepage_range(file, addr, len))
 253                         return -EINVAL;
 254                 return addr;
 255         }
 256
 257         if (addr) {
 258                 addr = ALIGN(addr, huge_page_size(h));
 259                 vma = find_vma(mm, addr);
 260                 if (mmap_end - len >= addr &&
 261                     (!vma || addr + len <= vm_start_gap(vma)))
 262                         return addr;
 263         }
 264
 265         /*
 266          * Use mm->get_unmapped_area value as a hint to use topdown routine.
 267          * If architectures have special needs, they should define their own
 268          * version of hugetlb_get_unmapped_area.
 269          */
 270         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
 271                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 272                                 pgoff, flags);
 273         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 274                         pgoff, flags);
 275 }
 276
 277 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 278 static unsigned long
 279 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 280                           unsigned long len, unsigned long pgoff,
 281                           unsigned long flags)
 282 {
 283         return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
 284 }
 285 #endif
 286
 287 /*
 288  * Support for read() - Find the page attached to f_mapping and copy out the
 289  * data. Its *very* similar to generic_file_buffered_read(), we can't use that
 290  * since it has PAGE_SIZE assumptions.
 291  */
 292 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 293 {
 294         struct file *file = iocb->ki_filp;
 295         struct hstate *h = hstate_file(file);
 296         struct address_space *mapping = file->f_mapping;
 297         struct inode *inode = mapping->host;
 298         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 299         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 300         unsigned long end_index;
 301         loff_t isize;
 302         ssize_t retval = 0;
 303
 304         while (iov_iter_count(to)) {
 305                 struct page *page;
 306                 size_t nr, copied;
 307
 308                 /* nr is the maximum number of bytes to copy from this page */
 309                 nr = huge_page_size(h);
 310                 isize = i_size_read(inode);
 311                 if (!isize)
 312                         break;
 313                 end_index = (isize - 1) >> huge_page_shift(h);
 314                 if (index > end_index)
 315                         break;
 316                 if (index == end_index) {
 317                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 318                         if (nr <= offset)
 319                                 break;
 320                 }
 321                 nr = nr - offset;
 322
 323                 /* Find the page */
 324                 page = find_lock_page(mapping, index);
 325                 if (unlikely(page == NULL)) {
 326                         /*
 327                          * We have a HOLE, zero out the user-buffer for the
 328                          * length of the hole or request.
 329                          */
 330                         copied = iov_iter_zero(nr, to);
 331                 } else {
 332                         unlock_page(page);
 333
 334                         /*
 335                          * We have the page, copy it to user space buffer.
 336                          */
 337                         copied = copy_page_to_iter(page, offset, nr, to);
 338                         put_page(page);
 339                 }
 340                 offset += copied;
 341                 retval += copied;
 342                 if (copied != nr && iov_iter_count(to)) {
 343                         if (!retval)
 344                                 retval = -EFAULT;
 345                         break;
 346                 }
 347                 index += offset >> huge_page_shift(h);
 348                 offset &= ~huge_page_mask(h);
 349         }
 350         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 351         return retval;
 352 }
 353
 354 static int hugetlbfs_write_begin(struct file *file,
 355                         struct address_space *mapping,
 356                         loff_t pos, unsigned len,
 357                         struct page **pagep, void **fsdata)
 358 {
 359         return -EINVAL;
 360 }
 361
 362 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 363                         loff_t pos, unsigned len, unsigned copied,
 364                         struct page *page, void *fsdata)
 365 {
 366         BUG();
 367         return -EINVAL;
 368 }
 369
 370 static void remove_huge_page(struct page *page)
 371 {
 372         ClearPageDirty(page);
 373         ClearPageUptodate(page);
 374         delete_from_page_cache(page);
 375 }
 376
 377 static void
 378 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
 379                       zap_flags_t zap_flags)
 380 {
 381         struct vm_area_struct *vma;
 382
 383         /*
 384          * end == 0 indicates that the entire range after start should be
 385          * unmapped.  Note, end is exclusive, whereas the interval tree takes
 386          * an inclusive "last".
 387          */
 388         vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
 389                 unsigned long v_offset;
 390                 unsigned long v_end;
 391
 392                 /*
 393                  * Can the expression below overflow on 32-bit arches?
 394                  * No, because the interval tree returns us only those vmas
 395                  * which overlap the truncated area starting at pgoff,
 396                  * and no vma on a 32-bit arch can span beyond the 4GB.
 397                  */
 398                 if (vma->vm_pgoff < start)
 399                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 400                 else
 401                         v_offset = 0;
 402
 403                 if (!end)
 404                         v_end = vma->vm_end;
 405                 else {
 406                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 407                                                         + vma->vm_start;
 408                         if (v_end > vma->vm_end)
 409                                 v_end = vma->vm_end;
 410                 }
 411
 412                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 413                                      NULL, zap_flags);
 414         }
 415 }
 416
 417 /*
 418  * remove_inode_hugepages handles two distinct cases: truncation and hole
 419  * punch.  There are subtle differences in operation for each case.
 420  *
 421  * truncation is indicated by end of range being LLONG_MAX
 422  *      In this case, we first scan the range and release found pages.
 423  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
 424  *      maps and global counts.  Page faults can not race with truncation
 425  *      in this routine.  hugetlb_no_page() holds i_mmap_rwsem and prevents
 426  *      page faults in the truncated range by checking i_size.  i_size is
 427  *      modified while holding i_mmap_rwsem.
 428  * hole punch is indicated if end is not LLONG_MAX
 429  *      In the hole punch case we scan the range and release found pages.
 430  *      Only when releasing a page is the associated region/reserve map
 431  *      deleted.  The region/reserve map for ranges without associated
 432  *      pages are not modified.  Page faults can race with hole punch.
 433  *      This is indicated if we find a mapped page.
 434  * Note: If the passed end of range value is beyond the end of file, but
 435  * not LLONG_MAX this routine still performs a hole punch operation.
 436  */
 437 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 438                                    loff_t lend)
 439 {
 440         struct hstate *h = hstate_inode(inode);
 441         struct address_space *mapping = &inode->i_data;
 442         const pgoff_t start = lstart >> huge_page_shift(h);
 443         const pgoff_t end = lend >> huge_page_shift(h);
 444         struct folio_batch fbatch;
 445         pgoff_t next, index;
 446         int i, freed = 0;
 447         bool truncate_op = (lend == LLONG_MAX);
 448
 449         folio_batch_init(&fbatch);
 450         next = start;
 451         while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
 452                 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
 453                         struct folio *folio = fbatch.folios[i];
 454                         u32 hash = 0;
 455
 456                         index = folio->index;
 457                         if (!truncate_op) {
 458                                 /*
 459                                  * Only need to hold the fault mutex in the
 460                                  * hole punch case.  This prevents races with
 461                                  * page faults.  Races are not possible in the
 462                                  * case of truncation.
 463                                  */
 464                                 hash = hugetlb_fault_mutex_hash(mapping, index);
 465                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 466                         }
 467
 468                         /*
 469                          * If folio is mapped, it was faulted in after being
 470                          * unmapped in caller.  Unmap (again) now after taking
 471                          * the fault mutex.  The mutex will prevent faults
 472                          * until we finish removing the folio.
 473                          *
 474                          * This race can only happen in the hole punch case.
 475                          * Getting here in a truncate operation is a bug.
 476                          */
 477                         if (unlikely(folio_mapped(folio))) {
 478                                 BUG_ON(truncate_op);
 479
 480                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 481                                 i_mmap_lock_write(mapping);
 482                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 483                                 hugetlb_vmdelete_list(&mapping->i_mmap,
 484                                         index * pages_per_huge_page(h),
 485                                         (index + 1) * pages_per_huge_page(h),
 486                                         ZAP_FLAG_DROP_MARKER);
 487                                 i_mmap_unlock_write(mapping);
 488                         }
 489
 490                         folio_lock(folio);
 491                         /*
 492                          * We must free the huge page and remove from page
 493                          * cache (remove_huge_page) BEFORE removing the
 494                          * region/reserve map (hugetlb_unreserve_pages).  In
 495                          * rare out of memory conditions, removal of the
 496                          * region/reserve map could fail. Correspondingly,
 497                          * the subpool and global reserve usage count can need
 498                          * to be adjusted.
 499                          */
 500                         VM_BUG_ON(HPageRestoreReserve(&folio->page));
 501                         remove_huge_page(&folio->page);
 502                         freed++;
 503                         if (!truncate_op) {
 504                                 if (unlikely(hugetlb_unreserve_pages(inode,
 505                                                         index, index + 1, 1)))
 506                                         hugetlb_fix_reserve_counts(inode);
 507                         }
 508
 509                         folio_unlock(folio);
 510                         if (!truncate_op)
 511                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 512                 }
 513                 folio_batch_release(&fbatch);
 514                 cond_resched();
 515         }
 516
 517         if (truncate_op)
 518                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 519 }
 520
 521 static void hugetlbfs_evict_inode(struct inode *inode)
 522 {
 523         struct resv_map *resv_map;
 524
 525         remove_inode_hugepages(inode, 0, LLONG_MAX);
 526
 527         /*
 528          * Get the resv_map from the address space embedded in the inode.
 529          * This is the address space which points to any resv_map allocated
 530          * at inode creation time.  If this is a device special inode,
 531          * i_mapping may not point to the original address space.
 532          */
 533         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
 534         /* Only regular and link inodes have associated reserve maps */
 535         if (resv_map)
 536                 resv_map_release(&resv_map->refs);
 537         clear_inode(inode);
 538 }
 539
 540 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 541 {
 542         pgoff_t pgoff;
 543         struct address_space *mapping = inode->i_mapping;
 544         struct hstate *h = hstate_inode(inode);
 545
 546         BUG_ON(offset & ~huge_page_mask(h));
 547         pgoff = offset >> PAGE_SHIFT;
 548
 549         i_mmap_lock_write(mapping);
 550         i_size_write(inode, offset);
 551         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 552                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
 553                                       ZAP_FLAG_DROP_MARKER);
 554         i_mmap_unlock_write(mapping);
 555         remove_inode_hugepages(inode, offset, LLONG_MAX);
 556 }
 557
 558 static void hugetlbfs_zero_partial_page(struct hstate *h,
 559                                         struct address_space *mapping,
 560                                         loff_t start,
 561                                         loff_t end)
 562 {
 563         pgoff_t idx = start >> huge_page_shift(h);
 564         struct folio *folio;
 565
 566         folio = filemap_lock_folio(mapping, idx);
 567         if (!folio)
 568                 return;
 569
 570         start = start & ~huge_page_mask(h);
 571         end = end & ~huge_page_mask(h);
 572         if (!end)
 573                 end = huge_page_size(h);
 574
 575         folio_zero_segment(folio, (size_t)start, (size_t)end);
 576
 577         folio_unlock(folio);
 578         folio_put(folio);
 579 }
 580
 581 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 582 {
 583         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 584         struct address_space *mapping = inode->i_mapping;
 585         struct hstate *h = hstate_inode(inode);
 586         loff_t hpage_size = huge_page_size(h);
 587         loff_t hole_start, hole_end;
 588
 589         /*
 590          * hole_start and hole_end indicate the full pages within the hole.
 591          */
 592         hole_start = round_up(offset, hpage_size);
 593         hole_end = round_down(offset + len, hpage_size);
 594
 595         inode_lock(inode);
 596
 597         /* protected by i_rwsem */
 598         if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 599                 inode_unlock(inode);
 600                 return -EPERM;
 601         }
 602
 603         i_mmap_lock_write(mapping);
 604
 605         /* If range starts before first full page, zero partial page. */
 606         if (offset < hole_start)
 607                 hugetlbfs_zero_partial_page(h, mapping,
 608                                 offset, min(offset + len, hole_start));
 609
 610         /* Unmap users of full pages in the hole. */
 611         if (hole_end > hole_start) {
 612                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 613                         hugetlb_vmdelete_list(&mapping->i_mmap,
 614                                               hole_start >> PAGE_SHIFT,
 615                                               hole_end >> PAGE_SHIFT, 0);
 616         }
 617
 618         /* If range extends beyond last full page, zero partial page. */
 619         if ((offset + len) > hole_end && (offset + len) > hole_start)
 620                 hugetlbfs_zero_partial_page(h, mapping,
 621                                 hole_end, offset + len);
 622
 623         i_mmap_unlock_write(mapping);
 624
 625         /* Remove full pages from the file. */
 626         if (hole_end > hole_start)
 627                 remove_inode_hugepages(inode, hole_start, hole_end);
 628
 629         inode_unlock(inode);
 630
 631         return 0;
 632 }
 633
 634 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 635                                 loff_t len)
 636 {
 637         struct inode *inode = file_inode(file);
 638         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 639         struct address_space *mapping = inode->i_mapping;
 640         struct hstate *h = hstate_inode(inode);
 641         struct vm_area_struct pseudo_vma;
 642         struct mm_struct *mm = current->mm;
 643         loff_t hpage_size = huge_page_size(h);
 644         unsigned long hpage_shift = huge_page_shift(h);
 645         pgoff_t start, index, end;
 646         int error;
 647         u32 hash;
 648
 649         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 650                 return -EOPNOTSUPP;
 651
 652         if (mode & FALLOC_FL_PUNCH_HOLE)
 653                 return hugetlbfs_punch_hole(inode, offset, len);
 654
 655         /*
 656          * Default preallocate case.
 657          * For this range, start is rounded down and end is rounded up
 658          * as well as being converted to page offsets.
 659          */
 660         start = offset >> hpage_shift;
 661         end = (offset + len + hpage_size - 1) >> hpage_shift;
 662
 663         inode_lock(inode);
 664
 665         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 666         error = inode_newsize_ok(inode, offset + len);
 667         if (error)
 668                 goto out;
 669
 670         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 671                 error = -EPERM;
 672                 goto out;
 673         }
 674
 675         /*
 676          * Initialize a pseudo vma as this is required by the huge page
 677          * allocation routines.  If NUMA is configured, use page index
 678          * as input to create an allocation policy.
 679          */
 680         vma_init(&pseudo_vma, mm);
 681         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 682         pseudo_vma.vm_file = file;
 683
 684         for (index = start; index < end; index++) {
 685                 /*
 686                  * This is supposed to be the vaddr where the page is being
 687                  * faulted in, but we have no vaddr here.
 688                  */
 689                 struct page *page;
 690                 unsigned long addr;
 691
 692                 cond_resched();
 693
 694                 /*
 695                  * fallocate(2) manpage permits EINTR; we may have been
 696                  * interrupted because we are using up too much memory.
 697                  */
 698                 if (signal_pending(current)) {
 699                         error = -EINTR;
 700                         break;
 701                 }
 702
 703                 /* Set numa allocation policy based on index */
 704                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 705
 706                 /* addr is the offset within the file (zero based) */
 707                 addr = index * hpage_size;
 708
 709                 /*
 710                  * fault mutex taken here, protects against fault path
 711                  * and hole punch.  inode_lock previously taken protects
 712                  * against truncation.
 713                  */
 714                 hash = hugetlb_fault_mutex_hash(mapping, index);
 715                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 716
 717                 /* See if already present in mapping to avoid alloc/free */
 718                 page = find_get_page(mapping, index);
 719                 if (page) {
 720                         put_page(page);
 721                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 722                         hugetlb_drop_vma_policy(&pseudo_vma);
 723                         continue;
 724                 }
 725
 726                 /*
 727                  * Allocate page without setting the avoid_reserve argument.
 728                  * There certainly are no reserves associated with the
 729                  * pseudo_vma.  However, there could be shared mappings with
 730                  * reserves for the file at the inode level.  If we fallocate
 731                  * pages in these areas, we need to consume the reserves
 732                  * to keep reservation accounting consistent.
 733                  */
 734                 page = alloc_huge_page(&pseudo_vma, addr, 0);
 735                 hugetlb_drop_vma_policy(&pseudo_vma);
 736                 if (IS_ERR(page)) {
 737                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 738                         error = PTR_ERR(page);
 739                         goto out;
 740                 }
 741                 clear_huge_page(page, addr, pages_per_huge_page(h));
 742                 __SetPageUptodate(page);
 743                 error = huge_add_to_page_cache(page, mapping, index);
 744                 if (unlikely(error)) {
 745                         restore_reserve_on_error(h, &pseudo_vma, addr, page);
 746                         put_page(page);
 747                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 748                         goto out;
 749                 }
 750
 751                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 752
 753                 SetHPageMigratable(page);
 754                 /*
 755                  * unlock_page because locked by huge_add_to_page_cache()
 756                  * put_page() due to reference from alloc_huge_page()
 757                  */
 758                 unlock_page(page);
 759                 put_page(page);
 760         }
 761
 762         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 763                 i_size_write(inode, offset + len);
 764         inode->i_ctime = current_time(inode);
 765 out:
 766         inode_unlock(inode);
 767         return error;
 768 }
 769
 770 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
 771                              struct dentry *dentry, struct iattr *attr)
 772 {
 773         struct inode *inode = d_inode(dentry);
 774         struct hstate *h = hstate_inode(inode);
 775         int error;
 776         unsigned int ia_valid = attr->ia_valid;
 777         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 778
 779         error = setattr_prepare(&init_user_ns, dentry, attr);
 780         if (error)
 781                 return error;
 782
 783         if (ia_valid & ATTR_SIZE) {
 784                 loff_t oldsize = inode->i_size;
 785                 loff_t newsize = attr->ia_size;
 786
 787                 if (newsize & ~huge_page_mask(h))
 788                         return -EINVAL;
 789                 /* protected by i_rwsem */
 790                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 791                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 792                         return -EPERM;
 793                 hugetlb_vmtruncate(inode, newsize);
 794         }
 795
 796         setattr_copy(&init_user_ns, inode, attr);
 797         mark_inode_dirty(inode);
 798         return 0;
 799 }
 800
 801 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 802                                         struct hugetlbfs_fs_context *ctx)
 803 {
 804         struct inode *inode;
 805
 806         inode = new_inode(sb);
 807         if (inode) {
 808                 inode->i_ino = get_next_ino();
 809                 inode->i_mode = S_IFDIR | ctx->mode;
 810                 inode->i_uid = ctx->uid;
 811                 inode->i_gid = ctx->gid;
 812                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 813                 inode->i_op = &hugetlbfs_dir_inode_operations;
 814                 inode->i_fop = &simple_dir_operations;
 815                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 816                 inc_nlink(inode);
 817                 lockdep_annotate_inode_mutex_key(inode);
 818         }
 819         return inode;
 820 }
 821
 822 /*
 823  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 824  * be taken from reclaim -- unlike regular filesystems. This needs an
 825  * annotation because huge_pmd_share() does an allocation under hugetlb's
 826  * i_mmap_rwsem.
 827  */
 828 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 829
 830 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 831                                         struct inode *dir,
 832                                         umode_t mode, dev_t dev)
 833 {
 834         struct inode *inode;
 835         struct resv_map *resv_map = NULL;
 836
 837         /*
 838          * Reserve maps are only needed for inodes that can have associated
 839          * page allocations.
 840          */
 841         if (S_ISREG(mode) || S_ISLNK(mode)) {
 842                 resv_map = resv_map_alloc();
 843                 if (!resv_map)
 844                         return NULL;
 845         }
 846
 847         inode = new_inode(sb);
 848         if (inode) {
 849                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 850
 851                 inode->i_ino = get_next_ino();
 852                 inode_init_owner(&init_user_ns, inode, dir, mode);
 853                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 854                                 &hugetlbfs_i_mmap_rwsem_key);
 855                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 856                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 857                 inode->i_mapping->private_data = resv_map;
 858                 info->seals = F_SEAL_SEAL;
 859                 switch (mode & S_IFMT) {
 860                 default:
 861                         init_special_inode(inode, mode, dev);
 862                         break;
 863                 case S_IFREG:
 864                         inode->i_op = &hugetlbfs_inode_operations;
 865                         inode->i_fop = &hugetlbfs_file_operations;
 866                         break;
 867                 case S_IFDIR:
 868                         inode->i_op = &hugetlbfs_dir_inode_operations;
 869                         inode->i_fop = &simple_dir_operations;
 870
 871                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 872                         inc_nlink(inode);
 873                         break;
 874                 case S_IFLNK:
 875                         inode->i_op = &page_symlink_inode_operations;
 876                         inode_nohighmem(inode);
 877                         break;
 878                 }
 879                 lockdep_annotate_inode_mutex_key(inode);
 880         } else {
 881                 if (resv_map)
 882                         kref_put(&resv_map->refs, resv_map_release);
 883         }
 884
 885         return inode;
 886 }
 887
 888 /*
 889  * File creation. Allocate an inode, and we're done..
 890  */
 891 static int do_hugetlbfs_mknod(struct inode *dir,
 892                         struct dentry *dentry,
 893                         umode_t mode,
 894                         dev_t dev,
 895                         bool tmpfile)
 896 {
 897         struct inode *inode;
 898         int error = -ENOSPC;
 899
 900         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 901         if (inode) {
 902                 dir->i_ctime = dir->i_mtime = current_time(dir);
 903                 if (tmpfile) {
 904                         d_tmpfile(dentry, inode);
 905                 } else {
 906                         d_instantiate(dentry, inode);
 907                         dget(dentry);/* Extra count - pin the dentry in core */
 908                 }
 909                 error = 0;
 910         }
 911         return error;
 912 }
 913
 914 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
 915                            struct dentry *dentry, umode_t mode, dev_t dev)
 916 {
 917         return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
 918 }
 919
 920 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
 921                            struct dentry *dentry, umode_t mode)
 922 {
 923         int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
 924                                      mode | S_IFDIR, 0);
 925         if (!retval)
 926                 inc_nlink(dir);
 927         return retval;
 928 }
 929
 930 static int hugetlbfs_create(struct user_namespace *mnt_userns,
 931                             struct inode *dir, struct dentry *dentry,
 932                             umode_t mode, bool excl)
 933 {
 934         return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
 935 }
 936
 937 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
 938                              struct inode *dir, struct dentry *dentry,
 939                              umode_t mode)
 940 {
 941         return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
 942 }
 943
 944 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
 945                              struct inode *dir, struct dentry *dentry,
 946                              const char *symname)
 947 {
 948         struct inode *inode;
 949         int error = -ENOSPC;
 950
 951         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 952         if (inode) {
 953                 int l = strlen(symname)+1;
 954                 error = page_symlink(inode, symname, l);
 955                 if (!error) {
 956                         d_instantiate(dentry, inode);
 957                         dget(dentry);
 958                 } else
 959                         iput(inode);
 960         }
 961         dir->i_ctime = dir->i_mtime = current_time(dir);
 962
 963         return error;
 964 }
 965
 966 #ifdef CONFIG_MIGRATION
 967 static int hugetlbfs_migrate_folio(struct address_space *mapping,
 968                                 struct folio *dst, struct folio *src,
 969                                 enum migrate_mode mode)
 970 {
 971         int rc;
 972
 973         rc = migrate_huge_page_move_mapping(mapping, dst, src);
 974         if (rc != MIGRATEPAGE_SUCCESS)
 975                 return rc;
 976
 977         if (hugetlb_page_subpool(&src->page)) {
 978                 hugetlb_set_page_subpool(&dst->page,
 979                                         hugetlb_page_subpool(&src->page));
 980                 hugetlb_set_page_subpool(&src->page, NULL);
 981         }
 982
 983         if (mode != MIGRATE_SYNC_NO_COPY)
 984                 folio_migrate_copy(dst, src);
 985         else
 986                 folio_migrate_flags(dst, src);
 987
 988         return MIGRATEPAGE_SUCCESS;
 989 }
 990 #else
 991 #define hugetlbfs_migrate_folio NULL
 992 #endif
 993
 994 static int hugetlbfs_error_remove_page(struct address_space *mapping,
 995                                 struct page *page)
 996 {
 997         struct inode *inode = mapping->host;
 998         pgoff_t index = page->index;
 999
1000         remove_huge_page(page);
1001         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
1002                 hugetlb_fix_reserve_counts(inode);
1003
1004         return 0;
1005 }
1006
1007 /*
1008  * Display the mount options in /proc/mounts.
1009  */
1010 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1011 {
1012         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1013         struct hugepage_subpool *spool = sbinfo->spool;
1014         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1015         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1016         char mod;
1017
1018         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1019                 seq_printf(m, ",uid=%u",
1020                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1021         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1022                 seq_printf(m, ",gid=%u",
1023                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1024         if (sbinfo->mode != 0755)
1025                 seq_printf(m, ",mode=%o", sbinfo->mode);
1026         if (sbinfo->max_inodes != -1)
1027                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1028
1029         hpage_size /= 1024;
1030         mod = 'K';
1031         if (hpage_size >= 1024) {
1032                 hpage_size /= 1024;
1033                 mod = 'M';
1034         }
1035         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1036         if (spool) {
1037                 if (spool->max_hpages != -1)
1038                         seq_printf(m, ",size=%llu",
1039                                    (unsigned long long)spool->max_hpages << hpage_shift);
1040                 if (spool->min_hpages != -1)
1041                         seq_printf(m, ",min_size=%llu",
1042                                    (unsigned long long)spool->min_hpages << hpage_shift);
1043         }
1044         return 0;
1045 }
1046
1047 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1048 {
1049         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1050         struct hstate *h = hstate_inode(d_inode(dentry));
1051
1052         buf->f_type = HUGETLBFS_MAGIC;
1053         buf->f_bsize = huge_page_size(h);
1054         if (sbinfo) {
1055                 spin_lock(&sbinfo->stat_lock);
1056                 /* If no limits set, just report 0 for max/free/used
1057                  * blocks, like simple_statfs() */
1058                 if (sbinfo->spool) {
1059                         long free_pages;
1060
1061                         spin_lock_irq(&sbinfo->spool->lock);
1062                         buf->f_blocks = sbinfo->spool->max_hpages;
1063                         free_pages = sbinfo->spool->max_hpages
1064                                 - sbinfo->spool->used_hpages;
1065                         buf->f_bavail = buf->f_bfree = free_pages;
1066                         spin_unlock_irq(&sbinfo->spool->lock);
1067                         buf->f_files = sbinfo->max_inodes;
1068                         buf->f_ffree = sbinfo->free_inodes;
1069                 }
1070                 spin_unlock(&sbinfo->stat_lock);
1071         }
1072         buf->f_namelen = NAME_MAX;
1073         return 0;
1074 }
1075
1076 static void hugetlbfs_put_super(struct super_block *sb)
1077 {
1078         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1079
1080         if (sbi) {
1081                 sb->s_fs_info = NULL;
1082
1083                 if (sbi->spool)
1084                         hugepage_put_subpool(sbi->spool);
1085
1086                 kfree(sbi);
1087         }
1088 }
1089
1090 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1091 {
1092         if (sbinfo->free_inodes >= 0) {
1093                 spin_lock(&sbinfo->stat_lock);
1094                 if (unlikely(!sbinfo->free_inodes)) {
1095                         spin_unlock(&sbinfo->stat_lock);
1096                         return 0;
1097                 }
1098                 sbinfo->free_inodes--;
1099                 spin_unlock(&sbinfo->stat_lock);
1100         }
1101
1102         return 1;
1103 }
1104
1105 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1106 {
1107         if (sbinfo->free_inodes >= 0) {
1108                 spin_lock(&sbinfo->stat_lock);
1109                 sbinfo->free_inodes++;
1110                 spin_unlock(&sbinfo->stat_lock);
1111         }
1112 }
1113
1114
1115 static struct kmem_cache *hugetlbfs_inode_cachep;
1116
1117 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1118 {
1119         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1120         struct hugetlbfs_inode_info *p;
1121
1122         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1123                 return NULL;
1124         p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1125         if (unlikely(!p)) {
1126                 hugetlbfs_inc_free_inodes(sbinfo);
1127                 return NULL;
1128         }
1129
1130         /*
1131          * Any time after allocation, hugetlbfs_destroy_inode can be called
1132          * for the inode.  mpol_free_shared_policy is unconditionally called
1133          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1134          * in case of a quick call to destroy.
1135          *
1136          * Note that the policy is initialized even if we are creating a
1137          * private inode.  This simplifies hugetlbfs_destroy_inode.
1138          */
1139         mpol_shared_policy_init(&p->policy, NULL);
1140
1141         return &p->vfs_inode;
1142 }
1143
1144 static void hugetlbfs_free_inode(struct inode *inode)
1145 {
1146         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1147 }
1148
1149 static void hugetlbfs_destroy_inode(struct inode *inode)
1150 {
1151         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1152         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1153 }
1154
1155 static const struct address_space_operations hugetlbfs_aops = {
1156         .write_begin    = hugetlbfs_write_begin,
1157         .write_end      = hugetlbfs_write_end,
1158         .dirty_folio    = noop_dirty_folio,
1159         .migrate_folio  = hugetlbfs_migrate_folio,
1160         .error_remove_page      = hugetlbfs_error_remove_page,
1161 };
1162
1163
1164 static void init_once(void *foo)
1165 {
1166         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1167
1168         inode_init_once(&ei->vfs_inode);
1169 }
1170
1171 const struct file_operations hugetlbfs_file_operations = {
1172         .read_iter              = hugetlbfs_read_iter,
1173         .mmap                   = hugetlbfs_file_mmap,
1174         .fsync                  = noop_fsync,
1175         .get_unmapped_area      = hugetlb_get_unmapped_area,
1176         .llseek                 = default_llseek,
1177         .fallocate              = hugetlbfs_fallocate,
1178 };
1179
1180 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1181         .create         = hugetlbfs_create,
1182         .lookup         = simple_lookup,
1183         .link           = simple_link,
1184         .unlink         = simple_unlink,
1185         .symlink        = hugetlbfs_symlink,
1186         .mkdir          = hugetlbfs_mkdir,
1187         .rmdir          = simple_rmdir,
1188         .mknod          = hugetlbfs_mknod,
1189         .rename         = simple_rename,
1190         .setattr        = hugetlbfs_setattr,
1191         .tmpfile        = hugetlbfs_tmpfile,
1192 };
1193
1194 static const struct inode_operations hugetlbfs_inode_operations = {
1195         .setattr        = hugetlbfs_setattr,
1196 };
1197
1198 static const struct super_operations hugetlbfs_ops = {
1199         .alloc_inode    = hugetlbfs_alloc_inode,
1200         .free_inode     = hugetlbfs_free_inode,
1201         .destroy_inode  = hugetlbfs_destroy_inode,
1202         .evict_inode    = hugetlbfs_evict_inode,
1203         .statfs         = hugetlbfs_statfs,
1204         .put_super      = hugetlbfs_put_super,
1205         .show_options   = hugetlbfs_show_options,
1206 };
1207
1208 /*
1209  * Convert size option passed from command line to number of huge pages
1210  * in the pool specified by hstate.  Size option could be in bytes
1211  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1212  */
1213 static long
1214 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1215                          enum hugetlbfs_size_type val_type)
1216 {
1217         if (val_type == NO_SIZE)
1218                 return -1;
1219
1220         if (val_type == SIZE_PERCENT) {
1221                 size_opt <<= huge_page_shift(h);
1222                 size_opt *= h->max_huge_pages;
1223                 do_div(size_opt, 100);
1224         }
1225
1226         size_opt >>= huge_page_shift(h);
1227         return size_opt;
1228 }
1229
1230 /*
1231  * Parse one mount parameter.
1232  */
1233 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1234 {
1235         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1236         struct fs_parse_result result;
1237         char *rest;
1238         unsigned long ps;
1239         int opt;
1240
1241         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1242         if (opt < 0)
1243                 return opt;
1244
1245         switch (opt) {
1246         case Opt_uid:
1247                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1248                 if (!uid_valid(ctx->uid))
1249                         goto bad_val;
1250                 return 0;
1251
1252         case Opt_gid:
1253                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1254                 if (!gid_valid(ctx->gid))
1255                         goto bad_val;
1256                 return 0;
1257
1258         case Opt_mode:
1259                 ctx->mode = result.uint_32 & 01777U;
1260                 return 0;
1261
1262         case Opt_size:
1263                 /* memparse() will accept a K/M/G without a digit */
1264                 if (!isdigit(param->string[0]))
1265                         goto bad_val;
1266                 ctx->max_size_opt = memparse(param->string, &rest);
1267                 ctx->max_val_type = SIZE_STD;
1268                 if (*rest == '%')
1269                         ctx->max_val_type = SIZE_PERCENT;
1270                 return 0;
1271
1272         case Opt_nr_inodes:
1273                 /* memparse() will accept a K/M/G without a digit */
1274                 if (!isdigit(param->string[0]))
1275                         goto bad_val;
1276                 ctx->nr_inodes = memparse(param->string, &rest);
1277                 return 0;
1278
1279         case Opt_pagesize:
1280                 ps = memparse(param->string, &rest);
1281                 ctx->hstate = size_to_hstate(ps);
1282                 if (!ctx->hstate) {
1283                         pr_err("Unsupported page size %lu MB\n", ps >> 20);
1284                         return -EINVAL;
1285                 }
1286                 return 0;
1287
1288         case Opt_min_size:
1289                 /* memparse() will accept a K/M/G without a digit */
1290                 if (!isdigit(param->string[0]))
1291                         goto bad_val;
1292                 ctx->min_size_opt = memparse(param->string, &rest);
1293                 ctx->min_val_type = SIZE_STD;
1294                 if (*rest == '%')
1295                         ctx->min_val_type = SIZE_PERCENT;
1296                 return 0;
1297
1298         default:
1299                 return -EINVAL;
1300         }
1301
1302 bad_val:
1303         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1304                       param->string, param->key);
1305 }
1306
1307 /*
1308  * Validate the parsed options.
1309  */
1310 static int hugetlbfs_validate(struct fs_context *fc)
1311 {
1312         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1313
1314         /*
1315          * Use huge page pool size (in hstate) to convert the size
1316          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1317          */
1318         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1319                                                    ctx->max_size_opt,
1320                                                    ctx->max_val_type);
1321         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1322                                                    ctx->min_size_opt,
1323                                                    ctx->min_val_type);
1324
1325         /*
1326          * If max_size was specified, then min_size must be smaller
1327          */
1328         if (ctx->max_val_type > NO_SIZE &&
1329             ctx->min_hpages > ctx->max_hpages) {
1330                 pr_err("Minimum size can not be greater than maximum size\n");
1331                 return -EINVAL;
1332         }
1333
1334         return 0;
1335 }
1336
1337 static int
1338 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1339 {
1340         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1341         struct hugetlbfs_sb_info *sbinfo;
1342
1343         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1344         if (!sbinfo)
1345                 return -ENOMEM;
1346         sb->s_fs_info = sbinfo;
1347         spin_lock_init(&sbinfo->stat_lock);
1348         sbinfo->hstate          = ctx->hstate;
1349         sbinfo->max_inodes      = ctx->nr_inodes;
1350         sbinfo->free_inodes     = ctx->nr_inodes;
1351         sbinfo->spool           = NULL;
1352         sbinfo->uid             = ctx->uid;
1353         sbinfo->gid             = ctx->gid;
1354         sbinfo->mode            = ctx->mode;
1355
1356         /*
1357          * Allocate and initialize subpool if maximum or minimum size is
1358          * specified.  Any needed reservations (for minimum size) are taken
1359          * taken when the subpool is created.
1360          */
1361         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1362                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1363                                                      ctx->max_hpages,
1364                                                      ctx->min_hpages);
1365                 if (!sbinfo->spool)
1366                         goto out_free;
1367         }
1368         sb->s_maxbytes = MAX_LFS_FILESIZE;
1369         sb->s_blocksize = huge_page_size(ctx->hstate);
1370         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1371         sb->s_magic = HUGETLBFS_MAGIC;
1372         sb->s_op = &hugetlbfs_ops;
1373         sb->s_time_gran = 1;
1374
1375         /*
1376          * Due to the special and limited functionality of hugetlbfs, it does
1377          * not work well as a stacking filesystem.
1378          */
1379         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1380         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1381         if (!sb->s_root)
1382                 goto out_free;
1383         return 0;
1384 out_free:
1385         kfree(sbinfo->spool);
1386         kfree(sbinfo);
1387         return -ENOMEM;
1388 }
1389
1390 static int hugetlbfs_get_tree(struct fs_context *fc)
1391 {
1392         int err = hugetlbfs_validate(fc);
1393         if (err)
1394                 return err;
1395         return get_tree_nodev(fc, hugetlbfs_fill_super);
1396 }
1397
1398 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1399 {
1400         kfree(fc->fs_private);
1401 }
1402
1403 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1404         .free           = hugetlbfs_fs_context_free,
1405         .parse_param    = hugetlbfs_parse_param,
1406         .get_tree       = hugetlbfs_get_tree,
1407 };
1408
1409 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1410 {
1411         struct hugetlbfs_fs_context *ctx;
1412
1413         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1414         if (!ctx)
1415                 return -ENOMEM;
1416
1417         ctx->max_hpages = -1; /* No limit on size by default */
1418         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1419         ctx->uid        = current_fsuid();
1420         ctx->gid        = current_fsgid();
1421         ctx->mode       = 0755;
1422         ctx->hstate     = &default_hstate;
1423         ctx->min_hpages = -1; /* No default minimum size */
1424         ctx->max_val_type = NO_SIZE;
1425         ctx->min_val_type = NO_SIZE;
1426         fc->fs_private = ctx;
1427         fc->ops = &hugetlbfs_fs_context_ops;
1428         return 0;
1429 }
1430
1431 static struct file_system_type hugetlbfs_fs_type = {
1432         .name                   = "hugetlbfs",
1433         .init_fs_context        = hugetlbfs_init_fs_context,
1434         .parameters             = hugetlb_fs_parameters,
1435         .kill_sb                = kill_litter_super,
1436 };
1437
1438 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1439
1440 static int can_do_hugetlb_shm(void)
1441 {
1442         kgid_t shm_group;
1443         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1444         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1445 }
1446
1447 static int get_hstate_idx(int page_size_log)
1448 {
1449         struct hstate *h = hstate_sizelog(page_size_log);
1450
1451         if (!h)
1452                 return -1;
1453         return hstate_index(h);
1454 }
1455
1456 /*
1457  * Note that size should be aligned to proper hugepage size in caller side,
1458  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1459  */
1460 struct file *hugetlb_file_setup(const char *name, size_t size,
1461                                 vm_flags_t acctflag, int creat_flags,
1462                                 int page_size_log)
1463 {
1464         struct inode *inode;
1465         struct vfsmount *mnt;
1466         int hstate_idx;
1467         struct file *file;
1468
1469         hstate_idx = get_hstate_idx(page_size_log);
1470         if (hstate_idx < 0)
1471                 return ERR_PTR(-ENODEV);
1472
1473         mnt = hugetlbfs_vfsmount[hstate_idx];
1474         if (!mnt)
1475                 return ERR_PTR(-ENOENT);
1476
1477         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1478                 struct ucounts *ucounts = current_ucounts();
1479
1480                 if (user_shm_lock(size, ucounts)) {
1481                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1482                                 current->comm, current->pid);
1483                         user_shm_unlock(size, ucounts);
1484                 }
1485                 return ERR_PTR(-EPERM);
1486         }
1487
1488         file = ERR_PTR(-ENOSPC);
1489         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1490         if (!inode)
1491                 goto out;
1492         if (creat_flags == HUGETLB_SHMFS_INODE)
1493                 inode->i_flags |= S_PRIVATE;
1494
1495         inode->i_size = size;
1496         clear_nlink(inode);
1497
1498         if (!hugetlb_reserve_pages(inode, 0,
1499                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1500                         acctflag))
1501                 file = ERR_PTR(-ENOMEM);
1502         else
1503                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1504                                         &hugetlbfs_file_operations);
1505         if (!IS_ERR(file))
1506                 return file;
1507
1508         iput(inode);
1509 out:
1510         return file;
1511 }
1512
1513 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1514 {
1515         struct fs_context *fc;
1516         struct vfsmount *mnt;
1517
1518         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1519         if (IS_ERR(fc)) {
1520                 mnt = ERR_CAST(fc);
1521         } else {
1522                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1523                 ctx->hstate = h;
1524                 mnt = fc_mount(fc);
1525                 put_fs_context(fc);
1526         }
1527         if (IS_ERR(mnt))
1528                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1529                        huge_page_size(h) >> 10);
1530         return mnt;
1531 }
1532
1533 static int __init init_hugetlbfs_fs(void)
1534 {
1535         struct vfsmount *mnt;
1536         struct hstate *h;
1537         int error;
1538         int i;
1539
1540         if (!hugepages_supported()) {
1541                 pr_info("disabling because there are no supported hugepage sizes\n");
1542                 return -ENOTSUPP;
1543         }
1544
1545         error = -ENOMEM;
1546         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1547                                         sizeof(struct hugetlbfs_inode_info),
1548                                         0, SLAB_ACCOUNT, init_once);
1549         if (hugetlbfs_inode_cachep == NULL)
1550                 goto out;
1551
1552         error = register_filesystem(&hugetlbfs_fs_type);
1553         if (error)
1554                 goto out_free;
1555
1556         /* default hstate mount is required */
1557         mnt = mount_one_hugetlbfs(&default_hstate);
1558         if (IS_ERR(mnt)) {
1559                 error = PTR_ERR(mnt);
1560                 goto out_unreg;
1561         }
1562         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1563
1564         /* other hstates are optional */
1565         i = 0;
1566         for_each_hstate(h) {
1567                 if (i == default_hstate_idx) {
1568                         i++;
1569                         continue;
1570                 }
1571
1572                 mnt = mount_one_hugetlbfs(h);
1573                 if (IS_ERR(mnt))
1574                         hugetlbfs_vfsmount[i] = NULL;
1575                 else
1576                         hugetlbfs_vfsmount[i] = mnt;
1577                 i++;
1578         }
1579
1580         return 0;
1581
1582  out_unreg:
1583         (void)unregister_filesystem(&hugetlbfs_fs_type);
1584  out_free:
1585         kmem_cache_destroy(hugetlbfs_inode_cachep);
1586  out:
1587         return error;
1588 }
1589 fs_initcall(init_hugetlbfs_fs)