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/falloc.h>
  15 #include <linux/fs.h>
  16 #include <linux/mount.h>
  17 #include <linux/file.h>
  18 #include <linux/kernel.h>
  19 #include <linux/writeback.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/highmem.h>
  22 #include <linux/init.h>
  23 #include <linux/string.h>
  24 #include <linux/capability.h>
  25 #include <linux/ctype.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/hugetlb.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/fs_parser.h>
  30 #include <linux/mman.h>
  31 #include <linux/slab.h>
  32 #include <linux/dnotify.h>
  33 #include <linux/statfs.h>
  34 #include <linux/security.h>
  35 #include <linux/magic.h>
  36 #include <linux/migrate.h>
  37 #include <linux/uio.h>
  38
  39 #include <linux/uaccess.h>
  40 #include <linux/sched/mm.h>
  41
  42 static const struct address_space_operations hugetlbfs_aops;
  43 const struct file_operations hugetlbfs_file_operations;
  44 static const struct inode_operations hugetlbfs_dir_inode_operations;
  45 static const struct inode_operations hugetlbfs_inode_operations;
  46
  47 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
  48
  49 struct hugetlbfs_fs_context {
  50         struct hstate           *hstate;
  51         unsigned long long      max_size_opt;
  52         unsigned long long      min_size_opt;
  53         long                    max_hpages;
  54         long                    nr_inodes;
  55         long                    min_hpages;
  56         enum hugetlbfs_size_type max_val_type;
  57         enum hugetlbfs_size_type min_val_type;
  58         kuid_t                  uid;
  59         kgid_t                  gid;
  60         umode_t                 mode;
  61 };
  62
  63 int sysctl_hugetlb_shm_group;
  64
  65 enum hugetlb_param {
  66         Opt_gid,
  67         Opt_min_size,
  68         Opt_mode,
  69         Opt_nr_inodes,
  70         Opt_pagesize,
  71         Opt_size,
  72         Opt_uid,
  73 };
  74
  75 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
  76         fsparam_u32   ("gid",           Opt_gid),
  77         fsparam_string("min_size",      Opt_min_size),
  78         fsparam_u32oct("mode",          Opt_mode),
  79         fsparam_string("nr_inodes",     Opt_nr_inodes),
  80         fsparam_string("pagesize",      Opt_pagesize),
  81         fsparam_string("size",          Opt_size),
  82         fsparam_u32   ("uid",           Opt_uid),
  83         {}
  84 };
  85
  86 #ifdef CONFIG_NUMA
  87 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  88                                         struct inode *inode, pgoff_t index)
  89 {
  90         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  91                                                         index);
  92 }
  93
  94 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  95 {
  96         mpol_cond_put(vma->vm_policy);
  97 }
  98 #else
  99 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
 100                                         struct inode *inode, pgoff_t index)
 101 {
 102 }
 103
 104 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
 105 {
 106 }
 107 #endif
 108
 109 /*
 110  * Mask used when checking the page offset value passed in via system
 111  * calls.  This value will be converted to a loff_t which is signed.
 112  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 113  * value.  The extra bit (- 1 in the shift value) is to take the sign
 114  * bit into account.
 115  */
 116 #define PGOFF_LOFFT_MAX \
 117         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 118
 119 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 120 {
 121         struct inode *inode = file_inode(file);
 122         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 123         loff_t len, vma_len;
 124         int ret;
 125         struct hstate *h = hstate_file(file);
 126
 127         /*
 128          * vma address alignment (but not the pgoff alignment) has
 129          * already been checked by prepare_hugepage_range.  If you add
 130          * any error returns here, do so after setting VM_HUGETLB, so
 131          * is_vm_hugetlb_page tests below unmap_region go the right
 132          * way when do_mmap unwinds (may be important on powerpc
 133          * and ia64).
 134          */
 135         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 136         vma->vm_ops = &hugetlb_vm_ops;
 137
 138         ret = seal_check_future_write(info->seals, vma);
 139         if (ret)
 140                 return ret;
 141
 142         /*
 143          * page based offset in vm_pgoff could be sufficiently large to
 144          * overflow a loff_t when converted to byte offset.  This can
 145          * only happen on architectures where sizeof(loff_t) ==
 146          * sizeof(unsigned long).  So, only check in those instances.
 147          */
 148         if (sizeof(unsigned long) == sizeof(loff_t)) {
 149                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 150                         return -EINVAL;
 151         }
 152
 153         /* must be huge page aligned */
 154         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 155                 return -EINVAL;
 156
 157         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 158         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 159         /* check for overflow */
 160         if (len < vma_len)
 161                 return -EINVAL;
 162
 163         inode_lock(inode);
 164         file_accessed(file);
 165
 166         ret = -ENOMEM;
 167         if (!hugetlb_reserve_pages(inode,
 168                                 vma->vm_pgoff >> huge_page_order(h),
 169                                 len >> huge_page_shift(h), vma,
 170                                 vma->vm_flags))
 171                 goto out;
 172
 173         ret = 0;
 174         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 175                 i_size_write(inode, len);
 176 out:
 177         inode_unlock(inode);
 178
 179         return ret;
 180 }
 181
 182 /*
 183  * Called under mmap_write_lock(mm).
 184  */
 185
 186 static unsigned long
 187 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
 188                 unsigned long len, unsigned long pgoff, unsigned long flags)
 189 {
 190         struct hstate *h = hstate_file(file);
 191         struct vm_unmapped_area_info info;
 192
 193         info.flags = 0;
 194         info.length = len;
 195         info.low_limit = current->mm->mmap_base;
 196         info.high_limit = arch_get_mmap_end(addr, len, flags);
 197         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 198         info.align_offset = 0;
 199         return vm_unmapped_area(&info);
 200 }
 201
 202 static unsigned long
 203 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
 204                 unsigned long len, unsigned long pgoff, unsigned long flags)
 205 {
 206         struct hstate *h = hstate_file(file);
 207         struct vm_unmapped_area_info info;
 208
 209         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 210         info.length = len;
 211         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
 212         info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
 213         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 214         info.align_offset = 0;
 215         addr = vm_unmapped_area(&info);
 216
 217         /*
 218          * A failed mmap() very likely causes application failure,
 219          * so fall back to the bottom-up function here. This scenario
 220          * can happen with large stack limits and large mmap()
 221          * allocations.
 222          */
 223         if (unlikely(offset_in_page(addr))) {
 224                 VM_BUG_ON(addr != -ENOMEM);
 225                 info.flags = 0;
 226                 info.low_limit = current->mm->mmap_base;
 227                 info.high_limit = arch_get_mmap_end(addr, len, flags);
 228                 addr = vm_unmapped_area(&info);
 229         }
 230
 231         return addr;
 232 }
 233
 234 unsigned long
 235 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 236                                   unsigned long len, unsigned long pgoff,
 237                                   unsigned long flags)
 238 {
 239         struct mm_struct *mm = current->mm;
 240         struct vm_area_struct *vma;
 241         struct hstate *h = hstate_file(file);
 242         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
 243
 244         if (len & ~huge_page_mask(h))
 245                 return -EINVAL;
 246         if (len > TASK_SIZE)
 247                 return -ENOMEM;
 248
 249         if (flags & MAP_FIXED) {
 250                 if (prepare_hugepage_range(file, addr, len))
 251                         return -EINVAL;
 252                 return addr;
 253         }
 254
 255         if (addr) {
 256                 addr = ALIGN(addr, huge_page_size(h));
 257                 vma = find_vma(mm, addr);
 258                 if (mmap_end - len >= addr &&
 259                     (!vma || addr + len <= vm_start_gap(vma)))
 260                         return addr;
 261         }
 262
 263         /*
 264          * Use mm->get_unmapped_area value as a hint to use topdown routine.
 265          * If architectures have special needs, they should define their own
 266          * version of hugetlb_get_unmapped_area.
 267          */
 268         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
 269                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 270                                 pgoff, flags);
 271         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 272                         pgoff, flags);
 273 }
 274
 275 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 276 static unsigned long
 277 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 278                           unsigned long len, unsigned long pgoff,
 279                           unsigned long flags)
 280 {
 281         return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
 282 }
 283 #endif
 284
 285 static size_t
 286 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 287                         struct iov_iter *to, unsigned long size)
 288 {
 289         size_t copied = 0;
 290         int i, chunksize;
 291
 292         /* Find which 4k chunk and offset with in that chunk */
 293         i = offset >> PAGE_SHIFT;
 294         offset = offset & ~PAGE_MASK;
 295
 296         while (size) {
 297                 size_t n;
 298                 chunksize = PAGE_SIZE;
 299                 if (offset)
 300                         chunksize -= offset;
 301                 if (chunksize > size)
 302                         chunksize = size;
 303                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 304                 copied += n;
 305                 if (n != chunksize)
 306                         return copied;
 307                 offset = 0;
 308                 size -= chunksize;
 309                 i++;
 310         }
 311         return copied;
 312 }
 313
 314 /*
 315  * Support for read() - Find the page attached to f_mapping and copy out the
 316  * data. This provides functionality similar to filemap_read().
 317  */
 318 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 319 {
 320         struct file *file = iocb->ki_filp;
 321         struct hstate *h = hstate_file(file);
 322         struct address_space *mapping = file->f_mapping;
 323         struct inode *inode = mapping->host;
 324         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 325         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 326         unsigned long end_index;
 327         loff_t isize;
 328         ssize_t retval = 0;
 329
 330         while (iov_iter_count(to)) {
 331                 struct page *page;
 332                 size_t nr, copied;
 333
 334                 /* nr is the maximum number of bytes to copy from this page */
 335                 nr = huge_page_size(h);
 336                 isize = i_size_read(inode);
 337                 if (!isize)
 338                         break;
 339                 end_index = (isize - 1) >> huge_page_shift(h);
 340                 if (index > end_index)
 341                         break;
 342                 if (index == end_index) {
 343                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 344                         if (nr <= offset)
 345                                 break;
 346                 }
 347                 nr = nr - offset;
 348
 349                 /* Find the page */
 350                 page = find_lock_page(mapping, index);
 351                 if (unlikely(page == NULL)) {
 352                         /*
 353                          * We have a HOLE, zero out the user-buffer for the
 354                          * length of the hole or request.
 355                          */
 356                         copied = iov_iter_zero(nr, to);
 357                 } else {
 358                         unlock_page(page);
 359
 360                         /*
 361                          * We have the page, copy it to user space buffer.
 362                          */
 363                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 364                         put_page(page);
 365                 }
 366                 offset += copied;
 367                 retval += copied;
 368                 if (copied != nr && iov_iter_count(to)) {
 369                         if (!retval)
 370                                 retval = -EFAULT;
 371                         break;
 372                 }
 373                 index += offset >> huge_page_shift(h);
 374                 offset &= ~huge_page_mask(h);
 375         }
 376         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 377         return retval;
 378 }
 379
 380 static int hugetlbfs_write_begin(struct file *file,
 381                         struct address_space *mapping,
 382                         loff_t pos, unsigned len,
 383                         struct page **pagep, void **fsdata)
 384 {
 385         return -EINVAL;
 386 }
 387
 388 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 389                         loff_t pos, unsigned len, unsigned copied,
 390                         struct page *page, void *fsdata)
 391 {
 392         BUG();
 393         return -EINVAL;
 394 }
 395
 396 static void remove_huge_page(struct page *page)
 397 {
 398         ClearPageDirty(page);
 399         ClearPageUptodate(page);
 400         delete_from_page_cache(page);
 401 }
 402
 403 static void
 404 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
 405                       zap_flags_t zap_flags)
 406 {
 407         struct vm_area_struct *vma;
 408
 409         /*
 410          * end == 0 indicates that the entire range after start should be
 411          * unmapped.  Note, end is exclusive, whereas the interval tree takes
 412          * an inclusive "last".
 413          */
 414         vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
 415                 unsigned long v_offset;
 416                 unsigned long v_end;
 417
 418                 /*
 419                  * Can the expression below overflow on 32-bit arches?
 420                  * No, because the interval tree returns us only those vmas
 421                  * which overlap the truncated area starting at pgoff,
 422                  * and no vma on a 32-bit arch can span beyond the 4GB.
 423                  */
 424                 if (vma->vm_pgoff < start)
 425                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 426                 else
 427                         v_offset = 0;
 428
 429                 if (!end)
 430                         v_end = vma->vm_end;
 431                 else {
 432                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 433                                                         + vma->vm_start;
 434                         if (v_end > vma->vm_end)
 435                                 v_end = vma->vm_end;
 436                 }
 437
 438                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 439                                      NULL, zap_flags);
 440         }
 441 }
 442
 443 /*
 444  * remove_inode_hugepages handles two distinct cases: truncation and hole
 445  * punch.  There are subtle differences in operation for each case.
 446  *
 447  * truncation is indicated by end of range being LLONG_MAX
 448  *      In this case, we first scan the range and release found pages.
 449  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
 450  *      maps and global counts.  Page faults can not race with truncation
 451  *      in this routine.  hugetlb_no_page() holds i_mmap_rwsem and prevents
 452  *      page faults in the truncated range by checking i_size.  i_size is
 453  *      modified while holding i_mmap_rwsem.
 454  * hole punch is indicated if end is not LLONG_MAX
 455  *      In the hole punch case we scan the range and release found pages.
 456  *      Only when releasing a page is the associated region/reserve map
 457  *      deleted.  The region/reserve map for ranges without associated
 458  *      pages are not modified.  Page faults can race with hole punch.
 459  *      This is indicated if we find a mapped page.
 460  * Note: If the passed end of range value is beyond the end of file, but
 461  * not LLONG_MAX this routine still performs a hole punch operation.
 462  */
 463 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 464                                    loff_t lend)
 465 {
 466         struct hstate *h = hstate_inode(inode);
 467         struct address_space *mapping = &inode->i_data;
 468         const pgoff_t start = lstart >> huge_page_shift(h);
 469         const pgoff_t end = lend >> huge_page_shift(h);
 470         struct folio_batch fbatch;
 471         pgoff_t next, index;
 472         int i, freed = 0;
 473         bool truncate_op = (lend == LLONG_MAX);
 474
 475         folio_batch_init(&fbatch);
 476         next = start;
 477         while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
 478                 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
 479                         struct folio *folio = fbatch.folios[i];
 480                         u32 hash = 0;
 481
 482                         index = folio->index;
 483                         if (!truncate_op) {
 484                                 /*
 485                                  * Only need to hold the fault mutex in the
 486                                  * hole punch case.  This prevents races with
 487                                  * page faults.  Races are not possible in the
 488                                  * case of truncation.
 489                                  */
 490                                 hash = hugetlb_fault_mutex_hash(mapping, index);
 491                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 492                         }
 493
 494                         /*
 495                          * If folio is mapped, it was faulted in after being
 496                          * unmapped in caller.  Unmap (again) now after taking
 497                          * the fault mutex.  The mutex will prevent faults
 498                          * until we finish removing the folio.
 499                          *
 500                          * This race can only happen in the hole punch case.
 501                          * Getting here in a truncate operation is a bug.
 502                          */
 503                         if (unlikely(folio_mapped(folio))) {
 504                                 BUG_ON(truncate_op);
 505
 506                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 507                                 i_mmap_lock_write(mapping);
 508                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 509                                 hugetlb_vmdelete_list(&mapping->i_mmap,
 510                                         index * pages_per_huge_page(h),
 511                                         (index + 1) * pages_per_huge_page(h),
 512                                         ZAP_FLAG_DROP_MARKER);
 513                                 i_mmap_unlock_write(mapping);
 514                         }
 515
 516                         folio_lock(folio);
 517                         /*
 518                          * We must free the huge page and remove from page
 519                          * cache (remove_huge_page) BEFORE removing the
 520                          * region/reserve map (hugetlb_unreserve_pages).  In
 521                          * rare out of memory conditions, removal of the
 522                          * region/reserve map could fail. Correspondingly,
 523                          * the subpool and global reserve usage count can need
 524                          * to be adjusted.
 525                          */
 526                         VM_BUG_ON(HPageRestoreReserve(&folio->page));
 527                         remove_huge_page(&folio->page);
 528                         freed++;
 529                         if (!truncate_op) {
 530                                 if (unlikely(hugetlb_unreserve_pages(inode,
 531                                                         index, index + 1, 1)))
 532                                         hugetlb_fix_reserve_counts(inode);
 533                         }
 534
 535                         folio_unlock(folio);
 536                         if (!truncate_op)
 537                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 538                 }
 539                 folio_batch_release(&fbatch);
 540                 cond_resched();
 541         }
 542
 543         if (truncate_op)
 544                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 545 }
 546
 547 static void hugetlbfs_evict_inode(struct inode *inode)
 548 {
 549         struct resv_map *resv_map;
 550
 551         remove_inode_hugepages(inode, 0, LLONG_MAX);
 552
 553         /*
 554          * Get the resv_map from the address space embedded in the inode.
 555          * This is the address space which points to any resv_map allocated
 556          * at inode creation time.  If this is a device special inode,
 557          * i_mapping may not point to the original address space.
 558          */
 559         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
 560         /* Only regular and link inodes have associated reserve maps */
 561         if (resv_map)
 562                 resv_map_release(&resv_map->refs);
 563         clear_inode(inode);
 564 }
 565
 566 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 567 {
 568         pgoff_t pgoff;
 569         struct address_space *mapping = inode->i_mapping;
 570         struct hstate *h = hstate_inode(inode);
 571
 572         BUG_ON(offset & ~huge_page_mask(h));
 573         pgoff = offset >> PAGE_SHIFT;
 574
 575         i_mmap_lock_write(mapping);
 576         i_size_write(inode, offset);
 577         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 578                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
 579                                       ZAP_FLAG_DROP_MARKER);
 580         i_mmap_unlock_write(mapping);
 581         remove_inode_hugepages(inode, offset, LLONG_MAX);
 582 }
 583
 584 static void hugetlbfs_zero_partial_page(struct hstate *h,
 585                                         struct address_space *mapping,
 586                                         loff_t start,
 587                                         loff_t end)
 588 {
 589         pgoff_t idx = start >> huge_page_shift(h);
 590         struct folio *folio;
 591
 592         folio = filemap_lock_folio(mapping, idx);
 593         if (!folio)
 594                 return;
 595
 596         start = start & ~huge_page_mask(h);
 597         end = end & ~huge_page_mask(h);
 598         if (!end)
 599                 end = huge_page_size(h);
 600
 601         folio_zero_segment(folio, (size_t)start, (size_t)end);
 602
 603         folio_unlock(folio);
 604         folio_put(folio);
 605 }
 606
 607 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 608 {
 609         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 610         struct address_space *mapping = inode->i_mapping;
 611         struct hstate *h = hstate_inode(inode);
 612         loff_t hpage_size = huge_page_size(h);
 613         loff_t hole_start, hole_end;
 614
 615         /*
 616          * hole_start and hole_end indicate the full pages within the hole.
 617          */
 618         hole_start = round_up(offset, hpage_size);
 619         hole_end = round_down(offset + len, hpage_size);
 620
 621         inode_lock(inode);
 622
 623         /* protected by i_rwsem */
 624         if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 625                 inode_unlock(inode);
 626                 return -EPERM;
 627         }
 628
 629         i_mmap_lock_write(mapping);
 630
 631         /* If range starts before first full page, zero partial page. */
 632         if (offset < hole_start)
 633                 hugetlbfs_zero_partial_page(h, mapping,
 634                                 offset, min(offset + len, hole_start));
 635
 636         /* Unmap users of full pages in the hole. */
 637         if (hole_end > hole_start) {
 638                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 639                         hugetlb_vmdelete_list(&mapping->i_mmap,
 640                                               hole_start >> PAGE_SHIFT,
 641                                               hole_end >> PAGE_SHIFT, 0);
 642         }
 643
 644         /* If range extends beyond last full page, zero partial page. */
 645         if ((offset + len) > hole_end && (offset + len) > hole_start)
 646                 hugetlbfs_zero_partial_page(h, mapping,
 647                                 hole_end, offset + len);
 648
 649         i_mmap_unlock_write(mapping);
 650
 651         /* Remove full pages from the file. */
 652         if (hole_end > hole_start)
 653                 remove_inode_hugepages(inode, hole_start, hole_end);
 654
 655         inode_unlock(inode);
 656
 657         return 0;
 658 }
 659
 660 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 661                                 loff_t len)
 662 {
 663         struct inode *inode = file_inode(file);
 664         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 665         struct address_space *mapping = inode->i_mapping;
 666         struct hstate *h = hstate_inode(inode);
 667         struct vm_area_struct pseudo_vma;
 668         struct mm_struct *mm = current->mm;
 669         loff_t hpage_size = huge_page_size(h);
 670         unsigned long hpage_shift = huge_page_shift(h);
 671         pgoff_t start, index, end;
 672         int error;
 673         u32 hash;
 674
 675         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 676                 return -EOPNOTSUPP;
 677
 678         if (mode & FALLOC_FL_PUNCH_HOLE)
 679                 return hugetlbfs_punch_hole(inode, offset, len);
 680
 681         /*
 682          * Default preallocate case.
 683          * For this range, start is rounded down and end is rounded up
 684          * as well as being converted to page offsets.
 685          */
 686         start = offset >> hpage_shift;
 687         end = (offset + len + hpage_size - 1) >> hpage_shift;
 688
 689         inode_lock(inode);
 690
 691         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 692         error = inode_newsize_ok(inode, offset + len);
 693         if (error)
 694                 goto out;
 695
 696         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 697                 error = -EPERM;
 698                 goto out;
 699         }
 700
 701         /*
 702          * Initialize a pseudo vma as this is required by the huge page
 703          * allocation routines.  If NUMA is configured, use page index
 704          * as input to create an allocation policy.
 705          */
 706         vma_init(&pseudo_vma, mm);
 707         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 708         pseudo_vma.vm_file = file;
 709
 710         for (index = start; index < end; index++) {
 711                 /*
 712                  * This is supposed to be the vaddr where the page is being
 713                  * faulted in, but we have no vaddr here.
 714                  */
 715                 struct page *page;
 716                 unsigned long addr;
 717
 718                 cond_resched();
 719
 720                 /*
 721                  * fallocate(2) manpage permits EINTR; we may have been
 722                  * interrupted because we are using up too much memory.
 723                  */
 724                 if (signal_pending(current)) {
 725                         error = -EINTR;
 726                         break;
 727                 }
 728
 729                 /* Set numa allocation policy based on index */
 730                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 731
 732                 /* addr is the offset within the file (zero based) */
 733                 addr = index * hpage_size;
 734
 735                 /*
 736                  * fault mutex taken here, protects against fault path
 737                  * and hole punch.  inode_lock previously taken protects
 738                  * against truncation.
 739                  */
 740                 hash = hugetlb_fault_mutex_hash(mapping, index);
 741                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 742
 743                 /* See if already present in mapping to avoid alloc/free */
 744                 page = find_get_page(mapping, index);
 745                 if (page) {
 746                         put_page(page);
 747                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 748                         hugetlb_drop_vma_policy(&pseudo_vma);
 749                         continue;
 750                 }
 751
 752                 /*
 753                  * Allocate page without setting the avoid_reserve argument.
 754                  * There certainly are no reserves associated with the
 755                  * pseudo_vma.  However, there could be shared mappings with
 756                  * reserves for the file at the inode level.  If we fallocate
 757                  * pages in these areas, we need to consume the reserves
 758                  * to keep reservation accounting consistent.
 759                  */
 760                 page = alloc_huge_page(&pseudo_vma, addr, 0);
 761                 hugetlb_drop_vma_policy(&pseudo_vma);
 762                 if (IS_ERR(page)) {
 763                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 764                         error = PTR_ERR(page);
 765                         goto out;
 766                 }
 767                 clear_huge_page(page, addr, pages_per_huge_page(h));
 768                 __SetPageUptodate(page);
 769                 error = huge_add_to_page_cache(page, mapping, index);
 770                 if (unlikely(error)) {
 771                         restore_reserve_on_error(h, &pseudo_vma, addr, page);
 772                         put_page(page);
 773                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 774                         goto out;
 775                 }
 776
 777                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 778
 779                 SetHPageMigratable(page);
 780                 /*
 781                  * unlock_page because locked by huge_add_to_page_cache()
 782                  * put_page() due to reference from alloc_huge_page()
 783                  */
 784                 unlock_page(page);
 785                 put_page(page);
 786         }
 787
 788         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 789                 i_size_write(inode, offset + len);
 790         inode->i_ctime = current_time(inode);
 791 out:
 792         inode_unlock(inode);
 793         return error;
 794 }
 795
 796 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
 797                              struct dentry *dentry, struct iattr *attr)
 798 {
 799         struct inode *inode = d_inode(dentry);
 800         struct hstate *h = hstate_inode(inode);
 801         int error;
 802         unsigned int ia_valid = attr->ia_valid;
 803         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 804
 805         error = setattr_prepare(&init_user_ns, dentry, attr);
 806         if (error)
 807                 return error;
 808
 809         if (ia_valid & ATTR_SIZE) {
 810                 loff_t oldsize = inode->i_size;
 811                 loff_t newsize = attr->ia_size;
 812
 813                 if (newsize & ~huge_page_mask(h))
 814                         return -EINVAL;
 815                 /* protected by i_rwsem */
 816                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 817                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 818                         return -EPERM;
 819                 hugetlb_vmtruncate(inode, newsize);
 820         }
 821
 822         setattr_copy(&init_user_ns, inode, attr);
 823         mark_inode_dirty(inode);
 824         return 0;
 825 }
 826
 827 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 828                                         struct hugetlbfs_fs_context *ctx)
 829 {
 830         struct inode *inode;
 831
 832         inode = new_inode(sb);
 833         if (inode) {
 834                 inode->i_ino = get_next_ino();
 835                 inode->i_mode = S_IFDIR | ctx->mode;
 836                 inode->i_uid = ctx->uid;
 837                 inode->i_gid = ctx->gid;
 838                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 839                 inode->i_op = &hugetlbfs_dir_inode_operations;
 840                 inode->i_fop = &simple_dir_operations;
 841                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 842                 inc_nlink(inode);
 843                 lockdep_annotate_inode_mutex_key(inode);
 844         }
 845         return inode;
 846 }
 847
 848 /*
 849  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 850  * be taken from reclaim -- unlike regular filesystems. This needs an
 851  * annotation because huge_pmd_share() does an allocation under hugetlb's
 852  * i_mmap_rwsem.
 853  */
 854 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 855
 856 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 857                                         struct inode *dir,
 858                                         umode_t mode, dev_t dev)
 859 {
 860         struct inode *inode;
 861         struct resv_map *resv_map = NULL;
 862
 863         /*
 864          * Reserve maps are only needed for inodes that can have associated
 865          * page allocations.
 866          */
 867         if (S_ISREG(mode) || S_ISLNK(mode)) {
 868                 resv_map = resv_map_alloc();
 869                 if (!resv_map)
 870                         return NULL;
 871         }
 872
 873         inode = new_inode(sb);
 874         if (inode) {
 875                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 876
 877                 inode->i_ino = get_next_ino();
 878                 inode_init_owner(&init_user_ns, inode, dir, mode);
 879                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 880                                 &hugetlbfs_i_mmap_rwsem_key);
 881                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 882                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 883                 inode->i_mapping->private_data = resv_map;
 884                 info->seals = F_SEAL_SEAL;
 885                 switch (mode & S_IFMT) {
 886                 default:
 887                         init_special_inode(inode, mode, dev);
 888                         break;
 889                 case S_IFREG:
 890                         inode->i_op = &hugetlbfs_inode_operations;
 891                         inode->i_fop = &hugetlbfs_file_operations;
 892                         break;
 893                 case S_IFDIR:
 894                         inode->i_op = &hugetlbfs_dir_inode_operations;
 895                         inode->i_fop = &simple_dir_operations;
 896
 897                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 898                         inc_nlink(inode);
 899                         break;
 900                 case S_IFLNK:
 901                         inode->i_op = &page_symlink_inode_operations;
 902                         inode_nohighmem(inode);
 903                         break;
 904                 }
 905                 lockdep_annotate_inode_mutex_key(inode);
 906         } else {
 907                 if (resv_map)
 908                         kref_put(&resv_map->refs, resv_map_release);
 909         }
 910
 911         return inode;
 912 }
 913
 914 /*
 915  * File creation. Allocate an inode, and we're done..
 916  */
 917 static int do_hugetlbfs_mknod(struct inode *dir,
 918                         struct dentry *dentry,
 919                         umode_t mode,
 920                         dev_t dev,
 921                         bool tmpfile)
 922 {
 923         struct inode *inode;
 924         int error = -ENOSPC;
 925
 926         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 927         if (inode) {
 928                 dir->i_ctime = dir->i_mtime = current_time(dir);
 929                 if (tmpfile) {
 930                         d_tmpfile(dentry, inode);
 931                 } else {
 932                         d_instantiate(dentry, inode);
 933                         dget(dentry);/* Extra count - pin the dentry in core */
 934                 }
 935                 error = 0;
 936         }
 937         return error;
 938 }
 939
 940 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
 941                            struct dentry *dentry, umode_t mode, dev_t dev)
 942 {
 943         return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
 944 }
 945
 946 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
 947                            struct dentry *dentry, umode_t mode)
 948 {
 949         int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
 950                                      mode | S_IFDIR, 0);
 951         if (!retval)
 952                 inc_nlink(dir);
 953         return retval;
 954 }
 955
 956 static int hugetlbfs_create(struct user_namespace *mnt_userns,
 957                             struct inode *dir, struct dentry *dentry,
 958                             umode_t mode, bool excl)
 959 {
 960         return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
 961 }
 962
 963 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
 964                              struct inode *dir, struct dentry *dentry,
 965                              umode_t mode)
 966 {
 967         return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
 968 }
 969
 970 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
 971                              struct inode *dir, struct dentry *dentry,
 972                              const char *symname)
 973 {
 974         struct inode *inode;
 975         int error = -ENOSPC;
 976
 977         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 978         if (inode) {
 979                 int l = strlen(symname)+1;
 980                 error = page_symlink(inode, symname, l);
 981                 if (!error) {
 982                         d_instantiate(dentry, inode);
 983                         dget(dentry);
 984                 } else
 985                         iput(inode);
 986         }
 987         dir->i_ctime = dir->i_mtime = current_time(dir);
 988
 989         return error;
 990 }
 991
 992 #ifdef CONFIG_MIGRATION
 993 static int hugetlbfs_migrate_folio(struct address_space *mapping,
 994                                 struct folio *dst, struct folio *src,
 995                                 enum migrate_mode mode)
 996 {
 997         int rc;
 998
 999         rc = migrate_huge_page_move_mapping(mapping, dst, src);
1000         if (rc != MIGRATEPAGE_SUCCESS)
1001                 return rc;
1002
1003         if (hugetlb_page_subpool(&src->page)) {
1004                 hugetlb_set_page_subpool(&dst->page,
1005                                         hugetlb_page_subpool(&src->page));
1006                 hugetlb_set_page_subpool(&src->page, NULL);
1007         }
1008
1009         if (mode != MIGRATE_SYNC_NO_COPY)
1010                 folio_migrate_copy(dst, src);
1011         else
1012                 folio_migrate_flags(dst, src);
1013
1014         return MIGRATEPAGE_SUCCESS;
1015 }
1016 #else
1017 #define hugetlbfs_migrate_folio NULL
1018 #endif
1019
1020 static int hugetlbfs_error_remove_page(struct address_space *mapping,
1021                                 struct page *page)
1022 {
1023         struct inode *inode = mapping->host;
1024         pgoff_t index = page->index;
1025
1026         remove_huge_page(page);
1027         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
1028                 hugetlb_fix_reserve_counts(inode);
1029
1030         return 0;
1031 }
1032
1033 /*
1034  * Display the mount options in /proc/mounts.
1035  */
1036 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1037 {
1038         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1039         struct hugepage_subpool *spool = sbinfo->spool;
1040         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1041         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1042         char mod;
1043
1044         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1045                 seq_printf(m, ",uid=%u",
1046                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1047         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1048                 seq_printf(m, ",gid=%u",
1049                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1050         if (sbinfo->mode != 0755)
1051                 seq_printf(m, ",mode=%o", sbinfo->mode);
1052         if (sbinfo->max_inodes != -1)
1053                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1054
1055         hpage_size /= 1024;
1056         mod = 'K';
1057         if (hpage_size >= 1024) {
1058                 hpage_size /= 1024;
1059                 mod = 'M';
1060         }
1061         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1062         if (spool) {
1063                 if (spool->max_hpages != -1)
1064                         seq_printf(m, ",size=%llu",
1065                                    (unsigned long long)spool->max_hpages << hpage_shift);
1066                 if (spool->min_hpages != -1)
1067                         seq_printf(m, ",min_size=%llu",
1068                                    (unsigned long long)spool->min_hpages << hpage_shift);
1069         }
1070         return 0;
1071 }
1072
1073 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1074 {
1075         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1076         struct hstate *h = hstate_inode(d_inode(dentry));
1077
1078         buf->f_type = HUGETLBFS_MAGIC;
1079         buf->f_bsize = huge_page_size(h);
1080         if (sbinfo) {
1081                 spin_lock(&sbinfo->stat_lock);
1082                 /* If no limits set, just report 0 or -1 for max/free/used
1083                  * blocks, like simple_statfs() */
1084                 if (sbinfo->spool) {
1085                         long free_pages;
1086
1087                         spin_lock_irq(&sbinfo->spool->lock);
1088                         buf->f_blocks = sbinfo->spool->max_hpages;
1089                         free_pages = sbinfo->spool->max_hpages
1090                                 - sbinfo->spool->used_hpages;
1091                         buf->f_bavail = buf->f_bfree = free_pages;
1092                         spin_unlock_irq(&sbinfo->spool->lock);
1093                         buf->f_files = sbinfo->max_inodes;
1094                         buf->f_ffree = sbinfo->free_inodes;
1095                 }
1096                 spin_unlock(&sbinfo->stat_lock);
1097         }
1098         buf->f_namelen = NAME_MAX;
1099         return 0;
1100 }
1101
1102 static void hugetlbfs_put_super(struct super_block *sb)
1103 {
1104         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1105
1106         if (sbi) {
1107                 sb->s_fs_info = NULL;
1108
1109                 if (sbi->spool)
1110                         hugepage_put_subpool(sbi->spool);
1111
1112                 kfree(sbi);
1113         }
1114 }
1115
1116 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1117 {
1118         if (sbinfo->free_inodes >= 0) {
1119                 spin_lock(&sbinfo->stat_lock);
1120                 if (unlikely(!sbinfo->free_inodes)) {
1121                         spin_unlock(&sbinfo->stat_lock);
1122                         return 0;
1123                 }
1124                 sbinfo->free_inodes--;
1125                 spin_unlock(&sbinfo->stat_lock);
1126         }
1127
1128         return 1;
1129 }
1130
1131 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1132 {
1133         if (sbinfo->free_inodes >= 0) {
1134                 spin_lock(&sbinfo->stat_lock);
1135                 sbinfo->free_inodes++;
1136                 spin_unlock(&sbinfo->stat_lock);
1137         }
1138 }
1139
1140
1141 static struct kmem_cache *hugetlbfs_inode_cachep;
1142
1143 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1144 {
1145         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1146         struct hugetlbfs_inode_info *p;
1147
1148         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1149                 return NULL;
1150         p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1151         if (unlikely(!p)) {
1152                 hugetlbfs_inc_free_inodes(sbinfo);
1153                 return NULL;
1154         }
1155
1156         /*
1157          * Any time after allocation, hugetlbfs_destroy_inode can be called
1158          * for the inode.  mpol_free_shared_policy is unconditionally called
1159          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1160          * in case of a quick call to destroy.
1161          *
1162          * Note that the policy is initialized even if we are creating a
1163          * private inode.  This simplifies hugetlbfs_destroy_inode.
1164          */
1165         mpol_shared_policy_init(&p->policy, NULL);
1166
1167         return &p->vfs_inode;
1168 }
1169
1170 static void hugetlbfs_free_inode(struct inode *inode)
1171 {
1172         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1173 }
1174
1175 static void hugetlbfs_destroy_inode(struct inode *inode)
1176 {
1177         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1178         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1179 }
1180
1181 static const struct address_space_operations hugetlbfs_aops = {
1182         .write_begin    = hugetlbfs_write_begin,
1183         .write_end      = hugetlbfs_write_end,
1184         .dirty_folio    = noop_dirty_folio,
1185         .migrate_folio  = hugetlbfs_migrate_folio,
1186         .error_remove_page      = hugetlbfs_error_remove_page,
1187 };
1188
1189
1190 static void init_once(void *foo)
1191 {
1192         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1193
1194         inode_init_once(&ei->vfs_inode);
1195 }
1196
1197 const struct file_operations hugetlbfs_file_operations = {
1198         .read_iter              = hugetlbfs_read_iter,
1199         .mmap                   = hugetlbfs_file_mmap,
1200         .fsync                  = noop_fsync,
1201         .get_unmapped_area      = hugetlb_get_unmapped_area,
1202         .llseek                 = default_llseek,
1203         .fallocate              = hugetlbfs_fallocate,
1204 };
1205
1206 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1207         .create         = hugetlbfs_create,
1208         .lookup         = simple_lookup,
1209         .link           = simple_link,
1210         .unlink         = simple_unlink,
1211         .symlink        = hugetlbfs_symlink,
1212         .mkdir          = hugetlbfs_mkdir,
1213         .rmdir          = simple_rmdir,
1214         .mknod          = hugetlbfs_mknod,
1215         .rename         = simple_rename,
1216         .setattr        = hugetlbfs_setattr,
1217         .tmpfile        = hugetlbfs_tmpfile,
1218 };
1219
1220 static const struct inode_operations hugetlbfs_inode_operations = {
1221         .setattr        = hugetlbfs_setattr,
1222 };
1223
1224 static const struct super_operations hugetlbfs_ops = {
1225         .alloc_inode    = hugetlbfs_alloc_inode,
1226         .free_inode     = hugetlbfs_free_inode,
1227         .destroy_inode  = hugetlbfs_destroy_inode,
1228         .evict_inode    = hugetlbfs_evict_inode,
1229         .statfs         = hugetlbfs_statfs,
1230         .put_super      = hugetlbfs_put_super,
1231         .show_options   = hugetlbfs_show_options,
1232 };
1233
1234 /*
1235  * Convert size option passed from command line to number of huge pages
1236  * in the pool specified by hstate.  Size option could be in bytes
1237  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1238  */
1239 static long
1240 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1241                          enum hugetlbfs_size_type val_type)
1242 {
1243         if (val_type == NO_SIZE)
1244                 return -1;
1245
1246         if (val_type == SIZE_PERCENT) {
1247                 size_opt <<= huge_page_shift(h);
1248                 size_opt *= h->max_huge_pages;
1249                 do_div(size_opt, 100);
1250         }
1251
1252         size_opt >>= huge_page_shift(h);
1253         return size_opt;
1254 }
1255
1256 /*
1257  * Parse one mount parameter.
1258  */
1259 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1260 {
1261         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1262         struct fs_parse_result result;
1263         char *rest;
1264         unsigned long ps;
1265         int opt;
1266
1267         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1268         if (opt < 0)
1269                 return opt;
1270
1271         switch (opt) {
1272         case Opt_uid:
1273                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1274                 if (!uid_valid(ctx->uid))
1275                         goto bad_val;
1276                 return 0;
1277
1278         case Opt_gid:
1279                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1280                 if (!gid_valid(ctx->gid))
1281                         goto bad_val;
1282                 return 0;
1283
1284         case Opt_mode:
1285                 ctx->mode = result.uint_32 & 01777U;
1286                 return 0;
1287
1288         case Opt_size:
1289                 /* memparse() will accept a K/M/G without a digit */
1290                 if (!isdigit(param->string[0]))
1291                         goto bad_val;
1292                 ctx->max_size_opt = memparse(param->string, &rest);
1293                 ctx->max_val_type = SIZE_STD;
1294                 if (*rest == '%')
1295                         ctx->max_val_type = SIZE_PERCENT;
1296                 return 0;
1297
1298         case Opt_nr_inodes:
1299                 /* memparse() will accept a K/M/G without a digit */
1300                 if (!isdigit(param->string[0]))
1301                         goto bad_val;
1302                 ctx->nr_inodes = memparse(param->string, &rest);
1303                 return 0;
1304
1305         case Opt_pagesize:
1306                 ps = memparse(param->string, &rest);
1307                 ctx->hstate = size_to_hstate(ps);
1308                 if (!ctx->hstate) {
1309                         pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1310                         return -EINVAL;
1311                 }
1312                 return 0;
1313
1314         case Opt_min_size:
1315                 /* memparse() will accept a K/M/G without a digit */
1316                 if (!isdigit(param->string[0]))
1317                         goto bad_val;
1318                 ctx->min_size_opt = memparse(param->string, &rest);
1319                 ctx->min_val_type = SIZE_STD;
1320                 if (*rest == '%')
1321                         ctx->min_val_type = SIZE_PERCENT;
1322                 return 0;
1323
1324         default:
1325                 return -EINVAL;
1326         }
1327
1328 bad_val:
1329         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1330                       param->string, param->key);
1331 }
1332
1333 /*
1334  * Validate the parsed options.
1335  */
1336 static int hugetlbfs_validate(struct fs_context *fc)
1337 {
1338         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1339
1340         /*
1341          * Use huge page pool size (in hstate) to convert the size
1342          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1343          */
1344         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1345                                                    ctx->max_size_opt,
1346                                                    ctx->max_val_type);
1347         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1348                                                    ctx->min_size_opt,
1349                                                    ctx->min_val_type);
1350
1351         /*
1352          * If max_size was specified, then min_size must be smaller
1353          */
1354         if (ctx->max_val_type > NO_SIZE &&
1355             ctx->min_hpages > ctx->max_hpages) {
1356                 pr_err("Minimum size can not be greater than maximum size\n");
1357                 return -EINVAL;
1358         }
1359
1360         return 0;
1361 }
1362
1363 static int
1364 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1365 {
1366         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1367         struct hugetlbfs_sb_info *sbinfo;
1368
1369         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1370         if (!sbinfo)
1371                 return -ENOMEM;
1372         sb->s_fs_info = sbinfo;
1373         spin_lock_init(&sbinfo->stat_lock);
1374         sbinfo->hstate          = ctx->hstate;
1375         sbinfo->max_inodes      = ctx->nr_inodes;
1376         sbinfo->free_inodes     = ctx->nr_inodes;
1377         sbinfo->spool           = NULL;
1378         sbinfo->uid             = ctx->uid;
1379         sbinfo->gid             = ctx->gid;
1380         sbinfo->mode            = ctx->mode;
1381
1382         /*
1383          * Allocate and initialize subpool if maximum or minimum size is
1384          * specified.  Any needed reservations (for minimum size) are taken
1385          * when the subpool is created.
1386          */
1387         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1388                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1389                                                      ctx->max_hpages,
1390                                                      ctx->min_hpages);
1391                 if (!sbinfo->spool)
1392                         goto out_free;
1393         }
1394         sb->s_maxbytes = MAX_LFS_FILESIZE;
1395         sb->s_blocksize = huge_page_size(ctx->hstate);
1396         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1397         sb->s_magic = HUGETLBFS_MAGIC;
1398         sb->s_op = &hugetlbfs_ops;
1399         sb->s_time_gran = 1;
1400
1401         /*
1402          * Due to the special and limited functionality of hugetlbfs, it does
1403          * not work well as a stacking filesystem.
1404          */
1405         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1406         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1407         if (!sb->s_root)
1408                 goto out_free;
1409         return 0;
1410 out_free:
1411         kfree(sbinfo->spool);
1412         kfree(sbinfo);
1413         return -ENOMEM;
1414 }
1415
1416 static int hugetlbfs_get_tree(struct fs_context *fc)
1417 {
1418         int err = hugetlbfs_validate(fc);
1419         if (err)
1420                 return err;
1421         return get_tree_nodev(fc, hugetlbfs_fill_super);
1422 }
1423
1424 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1425 {
1426         kfree(fc->fs_private);
1427 }
1428
1429 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1430         .free           = hugetlbfs_fs_context_free,
1431         .parse_param    = hugetlbfs_parse_param,
1432         .get_tree       = hugetlbfs_get_tree,
1433 };
1434
1435 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1436 {
1437         struct hugetlbfs_fs_context *ctx;
1438
1439         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1440         if (!ctx)
1441                 return -ENOMEM;
1442
1443         ctx->max_hpages = -1; /* No limit on size by default */
1444         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1445         ctx->uid        = current_fsuid();
1446         ctx->gid        = current_fsgid();
1447         ctx->mode       = 0755;
1448         ctx->hstate     = &default_hstate;
1449         ctx->min_hpages = -1; /* No default minimum size */
1450         ctx->max_val_type = NO_SIZE;
1451         ctx->min_val_type = NO_SIZE;
1452         fc->fs_private = ctx;
1453         fc->ops = &hugetlbfs_fs_context_ops;
1454         return 0;
1455 }
1456
1457 static struct file_system_type hugetlbfs_fs_type = {
1458         .name                   = "hugetlbfs",
1459         .init_fs_context        = hugetlbfs_init_fs_context,
1460         .parameters             = hugetlb_fs_parameters,
1461         .kill_sb                = kill_litter_super,
1462 };
1463
1464 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1465
1466 static int can_do_hugetlb_shm(void)
1467 {
1468         kgid_t shm_group;
1469         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1470         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1471 }
1472
1473 static int get_hstate_idx(int page_size_log)
1474 {
1475         struct hstate *h = hstate_sizelog(page_size_log);
1476
1477         if (!h)
1478                 return -1;
1479         return hstate_index(h);
1480 }
1481
1482 /*
1483  * Note that size should be aligned to proper hugepage size in caller side,
1484  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1485  */
1486 struct file *hugetlb_file_setup(const char *name, size_t size,
1487                                 vm_flags_t acctflag, int creat_flags,
1488                                 int page_size_log)
1489 {
1490         struct inode *inode;
1491         struct vfsmount *mnt;
1492         int hstate_idx;
1493         struct file *file;
1494
1495         hstate_idx = get_hstate_idx(page_size_log);
1496         if (hstate_idx < 0)
1497                 return ERR_PTR(-ENODEV);
1498
1499         mnt = hugetlbfs_vfsmount[hstate_idx];
1500         if (!mnt)
1501                 return ERR_PTR(-ENOENT);
1502
1503         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1504                 struct ucounts *ucounts = current_ucounts();
1505
1506                 if (user_shm_lock(size, ucounts)) {
1507                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1508                                 current->comm, current->pid);
1509                         user_shm_unlock(size, ucounts);
1510                 }
1511                 return ERR_PTR(-EPERM);
1512         }
1513
1514         file = ERR_PTR(-ENOSPC);
1515         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1516         if (!inode)
1517                 goto out;
1518         if (creat_flags == HUGETLB_SHMFS_INODE)
1519                 inode->i_flags |= S_PRIVATE;
1520
1521         inode->i_size = size;
1522         clear_nlink(inode);
1523
1524         if (!hugetlb_reserve_pages(inode, 0,
1525                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1526                         acctflag))
1527                 file = ERR_PTR(-ENOMEM);
1528         else
1529                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1530                                         &hugetlbfs_file_operations);
1531         if (!IS_ERR(file))
1532                 return file;
1533
1534         iput(inode);
1535 out:
1536         return file;
1537 }
1538
1539 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1540 {
1541         struct fs_context *fc;
1542         struct vfsmount *mnt;
1543
1544         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1545         if (IS_ERR(fc)) {
1546                 mnt = ERR_CAST(fc);
1547         } else {
1548                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1549                 ctx->hstate = h;
1550                 mnt = fc_mount(fc);
1551                 put_fs_context(fc);
1552         }
1553         if (IS_ERR(mnt))
1554                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1555                        huge_page_size(h) / SZ_1K);
1556         return mnt;
1557 }
1558
1559 static int __init init_hugetlbfs_fs(void)
1560 {
1561         struct vfsmount *mnt;
1562         struct hstate *h;
1563         int error;
1564         int i;
1565
1566         if (!hugepages_supported()) {
1567                 pr_info("disabling because there are no supported hugepage sizes\n");
1568                 return -ENOTSUPP;
1569         }
1570
1571         error = -ENOMEM;
1572         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1573                                         sizeof(struct hugetlbfs_inode_info),
1574                                         0, SLAB_ACCOUNT, init_once);
1575         if (hugetlbfs_inode_cachep == NULL)
1576                 goto out;
1577
1578         error = register_filesystem(&hugetlbfs_fs_type);
1579         if (error)
1580                 goto out_free;
1581
1582         /* default hstate mount is required */
1583         mnt = mount_one_hugetlbfs(&default_hstate);
1584         if (IS_ERR(mnt)) {
1585                 error = PTR_ERR(mnt);
1586                 goto out_unreg;
1587         }
1588         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1589
1590         /* other hstates are optional */
1591         i = 0;
1592         for_each_hstate(h) {
1593                 if (i == default_hstate_idx) {
1594                         i++;
1595                         continue;
1596                 }
1597
1598                 mnt = mount_one_hugetlbfs(h);
1599                 if (IS_ERR(mnt))
1600                         hugetlbfs_vfsmount[i] = NULL;
1601                 else
1602                         hugetlbfs_vfsmount[i] = mnt;
1603                 i++;
1604         }
1605
1606         return 0;
1607
1608  out_unreg:
1609         (void)unregister_filesystem(&hugetlbfs_fs_type);
1610  out_free:
1611         kmem_cache_destroy(hugetlbfs_inode_cachep);
1612  out:
1613         return error;
1614 }
1615 fs_initcall(init_hugetlbfs_fs)