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/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 struct hugetlbfs_config {
  49         struct hstate           *hstate;
  50         long                    max_hpages;
  51         long                    nr_inodes;
  52         long                    min_hpages;
  53         kuid_t                  uid;
  54         kgid_t                  gid;
  55         umode_t                 mode;
  56 };
  57
  58 int sysctl_hugetlb_shm_group;
  59
  60 enum {
  61         Opt_size, Opt_nr_inodes,
  62         Opt_mode, Opt_uid, Opt_gid,
  63         Opt_pagesize, Opt_min_size,
  64         Opt_err,
  65 };
  66
  67 static const match_table_t tokens = {
  68         {Opt_size,      "size=%s"},
  69         {Opt_nr_inodes, "nr_inodes=%s"},
  70         {Opt_mode,      "mode=%o"},
  71         {Opt_uid,       "uid=%u"},
  72         {Opt_gid,       "gid=%u"},
  73         {Opt_pagesize,  "pagesize=%s"},
  74         {Opt_min_size,  "min_size=%s"},
  75         {Opt_err,       NULL},
  76 };
  77
  78 #ifdef CONFIG_NUMA
  79 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  80                                         struct inode *inode, pgoff_t index)
  81 {
  82         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  83                                                         index);
  84 }
  85
  86 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  87 {
  88         mpol_cond_put(vma->vm_policy);
  89 }
  90 #else
  91 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  92                                         struct inode *inode, pgoff_t index)
  93 {
  94 }
  95
  96 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  97 {
  98 }
  99 #endif
 100
 101 static void huge_pagevec_release(struct pagevec *pvec)
 102 {
 103         int i;
 104
 105         for (i = 0; i < pagevec_count(pvec); ++i)
 106                 put_page(pvec->pages[i]);
 107
 108         pagevec_reinit(pvec);
 109 }
 110
 111 /*
 112  * Mask used when checking the page offset value passed in via system
 113  * calls.  This value will be converted to a loff_t which is signed.
 114  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 115  * value.  The extra bit (- 1 in the shift value) is to take the sign
 116  * bit into account.
 117  */
 118 #define PGOFF_LOFFT_MAX \
 119         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 120
 121 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 122 {
 123         struct inode *inode = file_inode(file);
 124         loff_t len, vma_len;
 125         int ret;
 126         struct hstate *h = hstate_file(file);
 127
 128         /*
 129          * vma address alignment (but not the pgoff alignment) has
 130          * already been checked by prepare_hugepage_range.  If you add
 131          * any error returns here, do so after setting VM_HUGETLB, so
 132          * is_vm_hugetlb_page tests below unmap_region go the right
 133          * way when do_mmap_pgoff unwinds (may be important on powerpc
 134          * and ia64).
 135          */
 136         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 137         vma->vm_ops = &hugetlb_vm_ops;
 138
 139         /*
 140          * page based offset in vm_pgoff could be sufficiently large to
 141          * overflow a loff_t when converted to byte offset.  This can
 142          * only happen on architectures where sizeof(loff_t) ==
 143          * sizeof(unsigned long).  So, only check in those instances.
 144          */
 145         if (sizeof(unsigned long) == sizeof(loff_t)) {
 146                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 147                         return -EINVAL;
 148         }
 149
 150         /* must be huge page aligned */
 151         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 152                 return -EINVAL;
 153
 154         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 155         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 156         /* check for overflow */
 157         if (len < vma_len)
 158                 return -EINVAL;
 159
 160         inode_lock(inode);
 161         file_accessed(file);
 162
 163         ret = -ENOMEM;
 164         if (hugetlb_reserve_pages(inode,
 165                                 vma->vm_pgoff >> huge_page_order(h),
 166                                 len >> huge_page_shift(h), vma,
 167                                 vma->vm_flags))
 168                 goto out;
 169
 170         ret = 0;
 171         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 172                 i_size_write(inode, len);
 173 out:
 174         inode_unlock(inode);
 175
 176         return ret;
 177 }
 178
 179 /*
 180  * Called under down_write(mmap_sem).
 181  */
 182
 183 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 184 static unsigned long
 185 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 186                 unsigned long len, unsigned long pgoff, unsigned long flags)
 187 {
 188         struct mm_struct *mm = current->mm;
 189         struct vm_area_struct *vma;
 190         struct hstate *h = hstate_file(file);
 191         struct vm_unmapped_area_info info;
 192
 193         if (len & ~huge_page_mask(h))
 194                 return -EINVAL;
 195         if (len > TASK_SIZE)
 196                 return -ENOMEM;
 197
 198         if (flags & MAP_FIXED) {
 199                 if (prepare_hugepage_range(file, addr, len))
 200                         return -EINVAL;
 201                 return addr;
 202         }
 203
 204         if (addr) {
 205                 addr = ALIGN(addr, huge_page_size(h));
 206                 vma = find_vma(mm, addr);
 207                 if (TASK_SIZE - len >= addr &&
 208                     (!vma || addr + len <= vm_start_gap(vma)))
 209                         return addr;
 210         }
 211
 212         info.flags = 0;
 213         info.length = len;
 214         info.low_limit = TASK_UNMAPPED_BASE;
 215         info.high_limit = TASK_SIZE;
 216         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 217         info.align_offset = 0;
 218         return vm_unmapped_area(&info);
 219 }
 220 #endif
 221
 222 static size_t
 223 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 224                         struct iov_iter *to, unsigned long size)
 225 {
 226         size_t copied = 0;
 227         int i, chunksize;
 228
 229         /* Find which 4k chunk and offset with in that chunk */
 230         i = offset >> PAGE_SHIFT;
 231         offset = offset & ~PAGE_MASK;
 232
 233         while (size) {
 234                 size_t n;
 235                 chunksize = PAGE_SIZE;
 236                 if (offset)
 237                         chunksize -= offset;
 238                 if (chunksize > size)
 239                         chunksize = size;
 240                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 241                 copied += n;
 242                 if (n != chunksize)
 243                         return copied;
 244                 offset = 0;
 245                 size -= chunksize;
 246                 i++;
 247         }
 248         return copied;
 249 }
 250
 251 /*
 252  * Support for read() - Find the page attached to f_mapping and copy out the
 253  * data. Its *very* similar to do_generic_mapping_read(), we can't use that
 254  * since it has PAGE_SIZE assumptions.
 255  */
 256 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 257 {
 258         struct file *file = iocb->ki_filp;
 259         struct hstate *h = hstate_file(file);
 260         struct address_space *mapping = file->f_mapping;
 261         struct inode *inode = mapping->host;
 262         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 263         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 264         unsigned long end_index;
 265         loff_t isize;
 266         ssize_t retval = 0;
 267
 268         while (iov_iter_count(to)) {
 269                 struct page *page;
 270                 size_t nr, copied;
 271
 272                 /* nr is the maximum number of bytes to copy from this page */
 273                 nr = huge_page_size(h);
 274                 isize = i_size_read(inode);
 275                 if (!isize)
 276                         break;
 277                 end_index = (isize - 1) >> huge_page_shift(h);
 278                 if (index > end_index)
 279                         break;
 280                 if (index == end_index) {
 281                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 282                         if (nr <= offset)
 283                                 break;
 284                 }
 285                 nr = nr - offset;
 286
 287                 /* Find the page */
 288                 page = find_lock_page(mapping, index);
 289                 if (unlikely(page == NULL)) {
 290                         /*
 291                          * We have a HOLE, zero out the user-buffer for the
 292                          * length of the hole or request.
 293                          */
 294                         copied = iov_iter_zero(nr, to);
 295                 } else {
 296                         unlock_page(page);
 297
 298                         /*
 299                          * We have the page, copy it to user space buffer.
 300                          */
 301                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 302                         put_page(page);
 303                 }
 304                 offset += copied;
 305                 retval += copied;
 306                 if (copied != nr && iov_iter_count(to)) {
 307                         if (!retval)
 308                                 retval = -EFAULT;
 309                         break;
 310                 }
 311                 index += offset >> huge_page_shift(h);
 312                 offset &= ~huge_page_mask(h);
 313         }
 314         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 315         return retval;
 316 }
 317
 318 static int hugetlbfs_write_begin(struct file *file,
 319                         struct address_space *mapping,
 320                         loff_t pos, unsigned len, unsigned flags,
 321                         struct page **pagep, void **fsdata)
 322 {
 323         return -EINVAL;
 324 }
 325
 326 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 327                         loff_t pos, unsigned len, unsigned copied,
 328                         struct page *page, void *fsdata)
 329 {
 330         BUG();
 331         return -EINVAL;
 332 }
 333
 334 static void remove_huge_page(struct page *page)
 335 {
 336         ClearPageDirty(page);
 337         ClearPageUptodate(page);
 338         delete_from_page_cache(page);
 339 }
 340
 341 static void
 342 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
 343 {
 344         struct vm_area_struct *vma;
 345
 346         /*
 347          * end == 0 indicates that the entire range after
 348          * start should be unmapped.
 349          */
 350         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 351                 unsigned long v_offset;
 352                 unsigned long v_end;
 353
 354                 /*
 355                  * Can the expression below overflow on 32-bit arches?
 356                  * No, because the interval tree returns us only those vmas
 357                  * which overlap the truncated area starting at pgoff,
 358                  * and no vma on a 32-bit arch can span beyond the 4GB.
 359                  */
 360                 if (vma->vm_pgoff < start)
 361                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 362                 else
 363                         v_offset = 0;
 364
 365                 if (!end)
 366                         v_end = vma->vm_end;
 367                 else {
 368                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 369                                                         + vma->vm_start;
 370                         if (v_end > vma->vm_end)
 371                                 v_end = vma->vm_end;
 372                 }
 373
 374                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 375                                                                         NULL);
 376         }
 377 }
 378
 379 /*
 380  * remove_inode_hugepages handles two distinct cases: truncation and hole
 381  * punch.  There are subtle differences in operation for each case.
 382  *
 383  * truncation is indicated by end of range being LLONG_MAX
 384  *      In this case, we first scan the range and release found pages.
 385  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
 386  *      maps and global counts.  Page faults can not race with truncation
 387  *      in this routine.  hugetlb_no_page() prevents page faults in the
 388  *      truncated range.  It checks i_size before allocation, and again after
 389  *      with the page table lock for the page held.  The same lock must be
 390  *      acquired to unmap a page.
 391  * hole punch is indicated if end is not LLONG_MAX
 392  *      In the hole punch case we scan the range and release found pages.
 393  *      Only when releasing a page is the associated region/reserv map
 394  *      deleted.  The region/reserv map for ranges without associated
 395  *      pages are not modified.  Page faults can race with hole punch.
 396  *      This is indicated if we find a mapped page.
 397  * Note: If the passed end of range value is beyond the end of file, but
 398  * not LLONG_MAX this routine still performs a hole punch operation.
 399  */
 400 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 401                                    loff_t lend)
 402 {
 403         struct hstate *h = hstate_inode(inode);
 404         struct address_space *mapping = &inode->i_data;
 405         const pgoff_t start = lstart >> huge_page_shift(h);
 406         const pgoff_t end = lend >> huge_page_shift(h);
 407         struct vm_area_struct pseudo_vma;
 408         struct pagevec pvec;
 409         pgoff_t next, index;
 410         int i, freed = 0;
 411         bool truncate_op = (lend == LLONG_MAX);
 412
 413         memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
 414         vma_init(&pseudo_vma, current->mm);
 415         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 416         pagevec_init(&pvec);
 417         next = start;
 418         while (next < end) {
 419                 /*
 420                  * When no more pages are found, we are done.
 421                  */
 422                 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
 423                         break;
 424
 425                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 426                         struct page *page = pvec.pages[i];
 427                         u32 hash;
 428
 429                         index = page->index;
 430                         hash = hugetlb_fault_mutex_hash(h, current->mm,
 431                                                         &pseudo_vma,
 432                                                         mapping, index, 0);
 433                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 434
 435                         /*
 436                          * If page is mapped, it was faulted in after being
 437                          * unmapped in caller.  Unmap (again) now after taking
 438                          * the fault mutex.  The mutex will prevent faults
 439                          * until we finish removing the page.
 440                          *
 441                          * This race can only happen in the hole punch case.
 442                          * Getting here in a truncate operation is a bug.
 443                          */
 444                         if (unlikely(page_mapped(page))) {
 445                                 BUG_ON(truncate_op);
 446
 447                                 i_mmap_lock_write(mapping);
 448                                 hugetlb_vmdelete_list(&mapping->i_mmap,
 449                                         index * pages_per_huge_page(h),
 450                                         (index + 1) * pages_per_huge_page(h));
 451                                 i_mmap_unlock_write(mapping);
 452                         }
 453
 454                         lock_page(page);
 455                         /*
 456                          * We must free the huge page and remove from page
 457                          * cache (remove_huge_page) BEFORE removing the
 458                          * region/reserve map (hugetlb_unreserve_pages).  In
 459                          * rare out of memory conditions, removal of the
 460                          * region/reserve map could fail. Correspondingly,
 461                          * the subpool and global reserve usage count can need
 462                          * to be adjusted.
 463                          */
 464                         VM_BUG_ON(PagePrivate(page));
 465                         remove_huge_page(page);
 466                         freed++;
 467                         if (!truncate_op) {
 468                                 if (unlikely(hugetlb_unreserve_pages(inode,
 469                                                         index, index + 1, 1)))
 470                                         hugetlb_fix_reserve_counts(inode);
 471                         }
 472
 473                         unlock_page(page);
 474                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 475                 }
 476                 huge_pagevec_release(&pvec);
 477                 cond_resched();
 478         }
 479
 480         if (truncate_op)
 481                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 482 }
 483
 484 static void hugetlbfs_evict_inode(struct inode *inode)
 485 {
 486         struct resv_map *resv_map;
 487
 488         remove_inode_hugepages(inode, 0, LLONG_MAX);
 489         resv_map = (struct resv_map *)inode->i_mapping->private_data;
 490         /* root inode doesn't have the resv_map, so we should check it */
 491         if (resv_map)
 492                 resv_map_release(&resv_map->refs);
 493         clear_inode(inode);
 494 }
 495
 496 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 497 {
 498         pgoff_t pgoff;
 499         struct address_space *mapping = inode->i_mapping;
 500         struct hstate *h = hstate_inode(inode);
 501
 502         BUG_ON(offset & ~huge_page_mask(h));
 503         pgoff = offset >> PAGE_SHIFT;
 504
 505         i_size_write(inode, offset);
 506         i_mmap_lock_write(mapping);
 507         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 508                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 509         i_mmap_unlock_write(mapping);
 510         remove_inode_hugepages(inode, offset, LLONG_MAX);
 511         return 0;
 512 }
 513
 514 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 515 {
 516         struct hstate *h = hstate_inode(inode);
 517         loff_t hpage_size = huge_page_size(h);
 518         loff_t hole_start, hole_end;
 519
 520         /*
 521          * For hole punch round up the beginning offset of the hole and
 522          * round down the end.
 523          */
 524         hole_start = round_up(offset, hpage_size);
 525         hole_end = round_down(offset + len, hpage_size);
 526
 527         if (hole_end > hole_start) {
 528                 struct address_space *mapping = inode->i_mapping;
 529                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 530
 531                 inode_lock(inode);
 532
 533                 /* protected by i_mutex */
 534                 if (info->seals & F_SEAL_WRITE) {
 535                         inode_unlock(inode);
 536                         return -EPERM;
 537                 }
 538
 539                 i_mmap_lock_write(mapping);
 540                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 541                         hugetlb_vmdelete_list(&mapping->i_mmap,
 542                                                 hole_start >> PAGE_SHIFT,
 543                                                 hole_end  >> PAGE_SHIFT);
 544                 i_mmap_unlock_write(mapping);
 545                 remove_inode_hugepages(inode, hole_start, hole_end);
 546                 inode_unlock(inode);
 547         }
 548
 549         return 0;
 550 }
 551
 552 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 553                                 loff_t len)
 554 {
 555         struct inode *inode = file_inode(file);
 556         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 557         struct address_space *mapping = inode->i_mapping;
 558         struct hstate *h = hstate_inode(inode);
 559         struct vm_area_struct pseudo_vma;
 560         struct mm_struct *mm = current->mm;
 561         loff_t hpage_size = huge_page_size(h);
 562         unsigned long hpage_shift = huge_page_shift(h);
 563         pgoff_t start, index, end;
 564         int error;
 565         u32 hash;
 566
 567         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 568                 return -EOPNOTSUPP;
 569
 570         if (mode & FALLOC_FL_PUNCH_HOLE)
 571                 return hugetlbfs_punch_hole(inode, offset, len);
 572
 573         /*
 574          * Default preallocate case.
 575          * For this range, start is rounded down and end is rounded up
 576          * as well as being converted to page offsets.
 577          */
 578         start = offset >> hpage_shift;
 579         end = (offset + len + hpage_size - 1) >> hpage_shift;
 580
 581         inode_lock(inode);
 582
 583         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 584         error = inode_newsize_ok(inode, offset + len);
 585         if (error)
 586                 goto out;
 587
 588         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 589                 error = -EPERM;
 590                 goto out;
 591         }
 592
 593         /*
 594          * Initialize a pseudo vma as this is required by the huge page
 595          * allocation routines.  If NUMA is configured, use page index
 596          * as input to create an allocation policy.
 597          */
 598         memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
 599         vma_init(&pseudo_vma, mm);
 600         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 601         pseudo_vma.vm_file = file;
 602
 603         for (index = start; index < end; index++) {
 604                 /*
 605                  * This is supposed to be the vaddr where the page is being
 606                  * faulted in, but we have no vaddr here.
 607                  */
 608                 struct page *page;
 609                 unsigned long addr;
 610                 int avoid_reserve = 0;
 611
 612                 cond_resched();
 613
 614                 /*
 615                  * fallocate(2) manpage permits EINTR; we may have been
 616                  * interrupted because we are using up too much memory.
 617                  */
 618                 if (signal_pending(current)) {
 619                         error = -EINTR;
 620                         break;
 621                 }
 622
 623                 /* Set numa allocation policy based on index */
 624                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 625
 626                 /* addr is the offset within the file (zero based) */
 627                 addr = index * hpage_size;
 628
 629                 /* mutex taken here, fault path and hole punch */
 630                 hash = hugetlb_fault_mutex_hash(h, mm, &pseudo_vma, mapping,
 631                                                 index, addr);
 632                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 633
 634                 /* See if already present in mapping to avoid alloc/free */
 635                 page = find_get_page(mapping, index);
 636                 if (page) {
 637                         put_page(page);
 638                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 639                         hugetlb_drop_vma_policy(&pseudo_vma);
 640                         continue;
 641                 }
 642
 643                 /* Allocate page and add to page cache */
 644                 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
 645                 hugetlb_drop_vma_policy(&pseudo_vma);
 646                 if (IS_ERR(page)) {
 647                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 648                         error = PTR_ERR(page);
 649                         goto out;
 650                 }
 651                 clear_huge_page(page, addr, pages_per_huge_page(h));
 652                 __SetPageUptodate(page);
 653                 error = huge_add_to_page_cache(page, mapping, index);
 654                 if (unlikely(error)) {
 655                         put_page(page);
 656                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 657                         goto out;
 658                 }
 659
 660                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 661
 662                 /*
 663                  * unlock_page because locked by add_to_page_cache()
 664                  * page_put due to reference from alloc_huge_page()
 665                  */
 666                 unlock_page(page);
 667                 put_page(page);
 668         }
 669
 670         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 671                 i_size_write(inode, offset + len);
 672         inode->i_ctime = current_time(inode);
 673 out:
 674         inode_unlock(inode);
 675         return error;
 676 }
 677
 678 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 679 {
 680         struct inode *inode = d_inode(dentry);
 681         struct hstate *h = hstate_inode(inode);
 682         int error;
 683         unsigned int ia_valid = attr->ia_valid;
 684         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 685
 686         BUG_ON(!inode);
 687
 688         error = setattr_prepare(dentry, attr);
 689         if (error)
 690                 return error;
 691
 692         if (ia_valid & ATTR_SIZE) {
 693                 loff_t oldsize = inode->i_size;
 694                 loff_t newsize = attr->ia_size;
 695
 696                 if (newsize & ~huge_page_mask(h))
 697                         return -EINVAL;
 698                 /* protected by i_mutex */
 699                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 700                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 701                         return -EPERM;
 702                 error = hugetlb_vmtruncate(inode, newsize);
 703                 if (error)
 704                         return error;
 705         }
 706
 707         setattr_copy(inode, attr);
 708         mark_inode_dirty(inode);
 709         return 0;
 710 }
 711
 712 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 713                                         struct hugetlbfs_config *config)
 714 {
 715         struct inode *inode;
 716
 717         inode = new_inode(sb);
 718         if (inode) {
 719                 inode->i_ino = get_next_ino();
 720                 inode->i_mode = S_IFDIR | config->mode;
 721                 inode->i_uid = config->uid;
 722                 inode->i_gid = config->gid;
 723                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 724                 inode->i_op = &hugetlbfs_dir_inode_operations;
 725                 inode->i_fop = &simple_dir_operations;
 726                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 727                 inc_nlink(inode);
 728                 lockdep_annotate_inode_mutex_key(inode);
 729         }
 730         return inode;
 731 }
 732
 733 /*
 734  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 735  * be taken from reclaim -- unlike regular filesystems. This needs an
 736  * annotation because huge_pmd_share() does an allocation under hugetlb's
 737  * i_mmap_rwsem.
 738  */
 739 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 740
 741 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 742                                         struct inode *dir,
 743                                         umode_t mode, dev_t dev)
 744 {
 745         struct inode *inode;
 746         struct resv_map *resv_map;
 747
 748         resv_map = resv_map_alloc();
 749         if (!resv_map)
 750                 return NULL;
 751
 752         inode = new_inode(sb);
 753         if (inode) {
 754                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 755
 756                 inode->i_ino = get_next_ino();
 757                 inode_init_owner(inode, dir, mode);
 758                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 759                                 &hugetlbfs_i_mmap_rwsem_key);
 760                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 761                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 762                 inode->i_mapping->private_data = resv_map;
 763                 info->seals = F_SEAL_SEAL;
 764                 switch (mode & S_IFMT) {
 765                 default:
 766                         init_special_inode(inode, mode, dev);
 767                         break;
 768                 case S_IFREG:
 769                         inode->i_op = &hugetlbfs_inode_operations;
 770                         inode->i_fop = &hugetlbfs_file_operations;
 771                         break;
 772                 case S_IFDIR:
 773                         inode->i_op = &hugetlbfs_dir_inode_operations;
 774                         inode->i_fop = &simple_dir_operations;
 775
 776                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 777                         inc_nlink(inode);
 778                         break;
 779                 case S_IFLNK:
 780                         inode->i_op = &page_symlink_inode_operations;
 781                         inode_nohighmem(inode);
 782                         break;
 783                 }
 784                 lockdep_annotate_inode_mutex_key(inode);
 785         } else
 786                 kref_put(&resv_map->refs, resv_map_release);
 787
 788         return inode;
 789 }
 790
 791 /*
 792  * File creation. Allocate an inode, and we're done..
 793  */
 794 static int hugetlbfs_mknod(struct inode *dir,
 795                         struct dentry *dentry, umode_t mode, dev_t dev)
 796 {
 797         struct inode *inode;
 798         int error = -ENOSPC;
 799
 800         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 801         if (inode) {
 802                 dir->i_ctime = dir->i_mtime = current_time(dir);
 803                 d_instantiate(dentry, inode);
 804                 dget(dentry);   /* Extra count - pin the dentry in core */
 805                 error = 0;
 806         }
 807         return error;
 808 }
 809
 810 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 811 {
 812         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 813         if (!retval)
 814                 inc_nlink(dir);
 815         return retval;
 816 }
 817
 818 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 819 {
 820         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 821 }
 822
 823 static int hugetlbfs_symlink(struct inode *dir,
 824                         struct dentry *dentry, const char *symname)
 825 {
 826         struct inode *inode;
 827         int error = -ENOSPC;
 828
 829         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 830         if (inode) {
 831                 int l = strlen(symname)+1;
 832                 error = page_symlink(inode, symname, l);
 833                 if (!error) {
 834                         d_instantiate(dentry, inode);
 835                         dget(dentry);
 836                 } else
 837                         iput(inode);
 838         }
 839         dir->i_ctime = dir->i_mtime = current_time(dir);
 840
 841         return error;
 842 }
 843
 844 /*
 845  * mark the head page dirty
 846  */
 847 static int hugetlbfs_set_page_dirty(struct page *page)
 848 {
 849         struct page *head = compound_head(page);
 850
 851         SetPageDirty(head);
 852         return 0;
 853 }
 854
 855 static int hugetlbfs_migrate_page(struct address_space *mapping,
 856                                 struct page *newpage, struct page *page,
 857                                 enum migrate_mode mode)
 858 {
 859         int rc;
 860
 861         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 862         if (rc != MIGRATEPAGE_SUCCESS)
 863                 return rc;
 864         if (mode != MIGRATE_SYNC_NO_COPY)
 865                 migrate_page_copy(newpage, page);
 866         else
 867                 migrate_page_states(newpage, page);
 868
 869         return MIGRATEPAGE_SUCCESS;
 870 }
 871
 872 static int hugetlbfs_error_remove_page(struct address_space *mapping,
 873                                 struct page *page)
 874 {
 875         struct inode *inode = mapping->host;
 876         pgoff_t index = page->index;
 877
 878         remove_huge_page(page);
 879         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
 880                 hugetlb_fix_reserve_counts(inode);
 881
 882         return 0;
 883 }
 884
 885 /*
 886  * Display the mount options in /proc/mounts.
 887  */
 888 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
 889 {
 890         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
 891         struct hugepage_subpool *spool = sbinfo->spool;
 892         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
 893         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
 894         char mod;
 895
 896         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
 897                 seq_printf(m, ",uid=%u",
 898                            from_kuid_munged(&init_user_ns, sbinfo->uid));
 899         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
 900                 seq_printf(m, ",gid=%u",
 901                            from_kgid_munged(&init_user_ns, sbinfo->gid));
 902         if (sbinfo->mode != 0755)
 903                 seq_printf(m, ",mode=%o", sbinfo->mode);
 904         if (sbinfo->max_inodes != -1)
 905                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
 906
 907         hpage_size /= 1024;
 908         mod = 'K';
 909         if (hpage_size >= 1024) {
 910                 hpage_size /= 1024;
 911                 mod = 'M';
 912         }
 913         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
 914         if (spool) {
 915                 if (spool->max_hpages != -1)
 916                         seq_printf(m, ",size=%llu",
 917                                    (unsigned long long)spool->max_hpages << hpage_shift);
 918                 if (spool->min_hpages != -1)
 919                         seq_printf(m, ",min_size=%llu",
 920                                    (unsigned long long)spool->min_hpages << hpage_shift);
 921         }
 922         return 0;
 923 }
 924
 925 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 926 {
 927         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 928         struct hstate *h = hstate_inode(d_inode(dentry));
 929
 930         buf->f_type = HUGETLBFS_MAGIC;
 931         buf->f_bsize = huge_page_size(h);
 932         if (sbinfo) {
 933                 spin_lock(&sbinfo->stat_lock);
 934                 /* If no limits set, just report 0 for max/free/used
 935                  * blocks, like simple_statfs() */
 936                 if (sbinfo->spool) {
 937                         long free_pages;
 938
 939                         spin_lock(&sbinfo->spool->lock);
 940                         buf->f_blocks = sbinfo->spool->max_hpages;
 941                         free_pages = sbinfo->spool->max_hpages
 942                                 - sbinfo->spool->used_hpages;
 943                         buf->f_bavail = buf->f_bfree = free_pages;
 944                         spin_unlock(&sbinfo->spool->lock);
 945                         buf->f_files = sbinfo->max_inodes;
 946                         buf->f_ffree = sbinfo->free_inodes;
 947                 }
 948                 spin_unlock(&sbinfo->stat_lock);
 949         }
 950         buf->f_namelen = NAME_MAX;
 951         return 0;
 952 }
 953
 954 static void hugetlbfs_put_super(struct super_block *sb)
 955 {
 956         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
 957
 958         if (sbi) {
 959                 sb->s_fs_info = NULL;
 960
 961                 if (sbi->spool)
 962                         hugepage_put_subpool(sbi->spool);
 963
 964                 kfree(sbi);
 965         }
 966 }
 967
 968 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 969 {
 970         if (sbinfo->free_inodes >= 0) {
 971                 spin_lock(&sbinfo->stat_lock);
 972                 if (unlikely(!sbinfo->free_inodes)) {
 973                         spin_unlock(&sbinfo->stat_lock);
 974                         return 0;
 975                 }
 976                 sbinfo->free_inodes--;
 977                 spin_unlock(&sbinfo->stat_lock);
 978         }
 979
 980         return 1;
 981 }
 982
 983 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 984 {
 985         if (sbinfo->free_inodes >= 0) {
 986                 spin_lock(&sbinfo->stat_lock);
 987                 sbinfo->free_inodes++;
 988                 spin_unlock(&sbinfo->stat_lock);
 989         }
 990 }
 991
 992
 993 static struct kmem_cache *hugetlbfs_inode_cachep;
 994
 995 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
 996 {
 997         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
 998         struct hugetlbfs_inode_info *p;
 999
1000         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1001                 return NULL;
1002         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1003         if (unlikely(!p)) {
1004                 hugetlbfs_inc_free_inodes(sbinfo);
1005                 return NULL;
1006         }
1007
1008         /*
1009          * Any time after allocation, hugetlbfs_destroy_inode can be called
1010          * for the inode.  mpol_free_shared_policy is unconditionally called
1011          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1012          * in case of a quick call to destroy.
1013          *
1014          * Note that the policy is initialized even if we are creating a
1015          * private inode.  This simplifies hugetlbfs_destroy_inode.
1016          */
1017         mpol_shared_policy_init(&p->policy, NULL);
1018
1019         return &p->vfs_inode;
1020 }
1021
1022 static void hugetlbfs_i_callback(struct rcu_head *head)
1023 {
1024         struct inode *inode = container_of(head, struct inode, i_rcu);
1025         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1026 }
1027
1028 static void hugetlbfs_destroy_inode(struct inode *inode)
1029 {
1030         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1031         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1032         call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
1033 }
1034
1035 static const struct address_space_operations hugetlbfs_aops = {
1036         .write_begin    = hugetlbfs_write_begin,
1037         .write_end      = hugetlbfs_write_end,
1038         .set_page_dirty = hugetlbfs_set_page_dirty,
1039         .migratepage    = hugetlbfs_migrate_page,
1040         .error_remove_page      = hugetlbfs_error_remove_page,
1041 };
1042
1043
1044 static void init_once(void *foo)
1045 {
1046         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1047
1048         inode_init_once(&ei->vfs_inode);
1049 }
1050
1051 const struct file_operations hugetlbfs_file_operations = {
1052         .read_iter              = hugetlbfs_read_iter,
1053         .mmap                   = hugetlbfs_file_mmap,
1054         .fsync                  = noop_fsync,
1055         .get_unmapped_area      = hugetlb_get_unmapped_area,
1056         .llseek                 = default_llseek,
1057         .fallocate              = hugetlbfs_fallocate,
1058 };
1059
1060 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1061         .create         = hugetlbfs_create,
1062         .lookup         = simple_lookup,
1063         .link           = simple_link,
1064         .unlink         = simple_unlink,
1065         .symlink        = hugetlbfs_symlink,
1066         .mkdir          = hugetlbfs_mkdir,
1067         .rmdir          = simple_rmdir,
1068         .mknod          = hugetlbfs_mknod,
1069         .rename         = simple_rename,
1070         .setattr        = hugetlbfs_setattr,
1071 };
1072
1073 static const struct inode_operations hugetlbfs_inode_operations = {
1074         .setattr        = hugetlbfs_setattr,
1075 };
1076
1077 static const struct super_operations hugetlbfs_ops = {
1078         .alloc_inode    = hugetlbfs_alloc_inode,
1079         .destroy_inode  = hugetlbfs_destroy_inode,
1080         .evict_inode    = hugetlbfs_evict_inode,
1081         .statfs         = hugetlbfs_statfs,
1082         .put_super      = hugetlbfs_put_super,
1083         .show_options   = hugetlbfs_show_options,
1084 };
1085
1086 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1087
1088 /*
1089  * Convert size option passed from command line to number of huge pages
1090  * in the pool specified by hstate.  Size option could be in bytes
1091  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1092  */
1093 static long
1094 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1095                          enum hugetlbfs_size_type val_type)
1096 {
1097         if (val_type == NO_SIZE)
1098                 return -1;
1099
1100         if (val_type == SIZE_PERCENT) {
1101                 size_opt <<= huge_page_shift(h);
1102                 size_opt *= h->max_huge_pages;
1103                 do_div(size_opt, 100);
1104         }
1105
1106         size_opt >>= huge_page_shift(h);
1107         return size_opt;
1108 }
1109
1110 static int
1111 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1112 {
1113         char *p, *rest;
1114         substring_t args[MAX_OPT_ARGS];
1115         int option;
1116         unsigned long long max_size_opt = 0, min_size_opt = 0;
1117         enum hugetlbfs_size_type max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1118
1119         if (!options)
1120                 return 0;
1121
1122         while ((p = strsep(&options, ",")) != NULL) {
1123                 int token;
1124                 if (!*p)
1125                         continue;
1126
1127                 token = match_token(p, tokens, args);
1128                 switch (token) {
1129                 case Opt_uid:
1130                         if (match_int(&args[0], &option))
1131                                 goto bad_val;
1132                         pconfig->uid = make_kuid(current_user_ns(), option);
1133                         if (!uid_valid(pconfig->uid))
1134                                 goto bad_val;
1135                         break;
1136
1137                 case Opt_gid:
1138                         if (match_int(&args[0], &option))
1139                                 goto bad_val;
1140                         pconfig->gid = make_kgid(current_user_ns(), option);
1141                         if (!gid_valid(pconfig->gid))
1142                                 goto bad_val;
1143                         break;
1144
1145                 case Opt_mode:
1146                         if (match_octal(&args[0], &option))
1147                                 goto bad_val;
1148                         pconfig->mode = option & 01777U;
1149                         break;
1150
1151                 case Opt_size: {
1152                         /* memparse() will accept a K/M/G without a digit */
1153                         if (!isdigit(*args[0].from))
1154                                 goto bad_val;
1155                         max_size_opt = memparse(args[0].from, &rest);
1156                         max_val_type = SIZE_STD;
1157                         if (*rest == '%')
1158                                 max_val_type = SIZE_PERCENT;
1159                         break;
1160                 }
1161
1162                 case Opt_nr_inodes:
1163                         /* memparse() will accept a K/M/G without a digit */
1164                         if (!isdigit(*args[0].from))
1165                                 goto bad_val;
1166                         pconfig->nr_inodes = memparse(args[0].from, &rest);
1167                         break;
1168
1169                 case Opt_pagesize: {
1170                         unsigned long ps;
1171                         ps = memparse(args[0].from, &rest);
1172                         pconfig->hstate = size_to_hstate(ps);
1173                         if (!pconfig->hstate) {
1174                                 pr_err("Unsupported page size %lu MB\n",
1175                                         ps >> 20);
1176                                 return -EINVAL;
1177                         }
1178                         break;
1179                 }
1180
1181                 case Opt_min_size: {
1182                         /* memparse() will accept a K/M/G without a digit */
1183                         if (!isdigit(*args[0].from))
1184                                 goto bad_val;
1185                         min_size_opt = memparse(args[0].from, &rest);
1186                         min_val_type = SIZE_STD;
1187                         if (*rest == '%')
1188                                 min_val_type = SIZE_PERCENT;
1189                         break;
1190                 }
1191
1192                 default:
1193                         pr_err("Bad mount option: \"%s\"\n", p);
1194                         return -EINVAL;
1195                         break;
1196                 }
1197         }
1198
1199         /*
1200          * Use huge page pool size (in hstate) to convert the size
1201          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1202          */
1203         pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1204                                                 max_size_opt, max_val_type);
1205         pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1206                                                 min_size_opt, min_val_type);
1207
1208         /*
1209          * If max_size was specified, then min_size must be smaller
1210          */
1211         if (max_val_type > NO_SIZE &&
1212             pconfig->min_hpages > pconfig->max_hpages) {
1213                 pr_err("minimum size can not be greater than maximum size\n");
1214                 return -EINVAL;
1215         }
1216
1217         return 0;
1218
1219 bad_val:
1220         pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1221         return -EINVAL;
1222 }
1223
1224 static int
1225 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1226 {
1227         int ret;
1228         struct hugetlbfs_config config;
1229         struct hugetlbfs_sb_info *sbinfo;
1230
1231         config.max_hpages = -1; /* No limit on size by default */
1232         config.nr_inodes = -1; /* No limit on number of inodes by default */
1233         config.uid = current_fsuid();
1234         config.gid = current_fsgid();
1235         config.mode = 0755;
1236         config.hstate = &default_hstate;
1237         config.min_hpages = -1; /* No default minimum size */
1238         ret = hugetlbfs_parse_options(data, &config);
1239         if (ret)
1240                 return ret;
1241
1242         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1243         if (!sbinfo)
1244                 return -ENOMEM;
1245         sb->s_fs_info = sbinfo;
1246         sbinfo->hstate = config.hstate;
1247         spin_lock_init(&sbinfo->stat_lock);
1248         sbinfo->max_inodes = config.nr_inodes;
1249         sbinfo->free_inodes = config.nr_inodes;
1250         sbinfo->spool = NULL;
1251         sbinfo->uid = config.uid;
1252         sbinfo->gid = config.gid;
1253         sbinfo->mode = config.mode;
1254
1255         /*
1256          * Allocate and initialize subpool if maximum or minimum size is
1257          * specified.  Any needed reservations (for minimim size) are taken
1258          * taken when the subpool is created.
1259          */
1260         if (config.max_hpages != -1 || config.min_hpages != -1) {
1261                 sbinfo->spool = hugepage_new_subpool(config.hstate,
1262                                                         config.max_hpages,
1263                                                         config.min_hpages);
1264                 if (!sbinfo->spool)
1265                         goto out_free;
1266         }
1267         sb->s_maxbytes = MAX_LFS_FILESIZE;
1268         sb->s_blocksize = huge_page_size(config.hstate);
1269         sb->s_blocksize_bits = huge_page_shift(config.hstate);
1270         sb->s_magic = HUGETLBFS_MAGIC;
1271         sb->s_op = &hugetlbfs_ops;
1272         sb->s_time_gran = 1;
1273         sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1274         if (!sb->s_root)
1275                 goto out_free;
1276         return 0;
1277 out_free:
1278         kfree(sbinfo->spool);
1279         kfree(sbinfo);
1280         return -ENOMEM;
1281 }
1282
1283 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1284         int flags, const char *dev_name, void *data)
1285 {
1286         return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1287 }
1288
1289 static struct file_system_type hugetlbfs_fs_type = {
1290         .name           = "hugetlbfs",
1291         .mount          = hugetlbfs_mount,
1292         .kill_sb        = kill_litter_super,
1293 };
1294
1295 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1296
1297 static int can_do_hugetlb_shm(void)
1298 {
1299         kgid_t shm_group;
1300         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1301         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1302 }
1303
1304 static int get_hstate_idx(int page_size_log)
1305 {
1306         struct hstate *h = hstate_sizelog(page_size_log);
1307
1308         if (!h)
1309                 return -1;
1310         return h - hstates;
1311 }
1312
1313 /*
1314  * Note that size should be aligned to proper hugepage size in caller side,
1315  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1316  */
1317 struct file *hugetlb_file_setup(const char *name, size_t size,
1318                                 vm_flags_t acctflag, struct user_struct **user,
1319                                 int creat_flags, int page_size_log)
1320 {
1321         struct inode *inode;
1322         struct vfsmount *mnt;
1323         int hstate_idx;
1324         struct file *file;
1325
1326         hstate_idx = get_hstate_idx(page_size_log);
1327         if (hstate_idx < 0)
1328                 return ERR_PTR(-ENODEV);
1329
1330         *user = NULL;
1331         mnt = hugetlbfs_vfsmount[hstate_idx];
1332         if (!mnt)
1333                 return ERR_PTR(-ENOENT);
1334
1335         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1336                 *user = current_user();
1337                 if (user_shm_lock(size, *user)) {
1338                         task_lock(current);
1339                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1340                                 current->comm, current->pid);
1341                         task_unlock(current);
1342                 } else {
1343                         *user = NULL;
1344                         return ERR_PTR(-EPERM);
1345                 }
1346         }
1347
1348         file = ERR_PTR(-ENOSPC);
1349         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1350         if (!inode)
1351                 goto out;
1352         if (creat_flags == HUGETLB_SHMFS_INODE)
1353                 inode->i_flags |= S_PRIVATE;
1354
1355         inode->i_size = size;
1356         clear_nlink(inode);
1357
1358         if (hugetlb_reserve_pages(inode, 0,
1359                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1360                         acctflag))
1361                 file = ERR_PTR(-ENOMEM);
1362         else
1363                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1364                                         &hugetlbfs_file_operations);
1365         if (!IS_ERR(file))
1366                 return file;
1367
1368         iput(inode);
1369 out:
1370         if (*user) {
1371                 user_shm_unlock(size, *user);
1372                 *user = NULL;
1373         }
1374         return file;
1375 }
1376
1377 static int __init init_hugetlbfs_fs(void)
1378 {
1379         struct hstate *h;
1380         int error;
1381         int i;
1382
1383         if (!hugepages_supported()) {
1384                 pr_info("disabling because there are no supported hugepage sizes\n");
1385                 return -ENOTSUPP;
1386         }
1387
1388         error = -ENOMEM;
1389         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1390                                         sizeof(struct hugetlbfs_inode_info),
1391                                         0, SLAB_ACCOUNT, init_once);
1392         if (hugetlbfs_inode_cachep == NULL)
1393                 goto out2;
1394
1395         error = register_filesystem(&hugetlbfs_fs_type);
1396         if (error)
1397                 goto out;
1398
1399         i = 0;
1400         for_each_hstate(h) {
1401                 char buf[50];
1402                 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1403
1404                 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1405                 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1406                                                         buf);
1407
1408                 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1409                         pr_err("Cannot mount internal hugetlbfs for "
1410                                 "page size %uK", ps_kb);
1411                         error = PTR_ERR(hugetlbfs_vfsmount[i]);
1412                         hugetlbfs_vfsmount[i] = NULL;
1413                 }
1414                 i++;
1415         }
1416         /* Non default hstates are optional */
1417         if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1418                 return 0;
1419
1420  out:
1421         kmem_cache_destroy(hugetlbfs_inode_cachep);
1422  out2:
1423         return error;
1424 }
1425 fs_initcall(init_hugetlbfs_fs)