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