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