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  */
   8
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/module.h>
  12 #include <linux/thread_info.h>
  13 #include <asm/current.h>
  14 #include <linux/sched.h>                /* remove ASAP */
  15 #include <linux/fs.h>
  16 #include <linux/mount.h>
  17 #include <linux/file.h>
  18 #include <linux/kernel.h>
  19 #include <linux/writeback.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/highmem.h>
  22 #include <linux/init.h>
  23 #include <linux/string.h>
  24 #include <linux/capability.h>
  25 #include <linux/ctype.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/hugetlb.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/parser.h>
  30 #include <linux/mman.h>
  31 #include <linux/slab.h>
  32 #include <linux/dnotify.h>
  33 #include <linux/statfs.h>
  34 #include <linux/security.h>
  35 #include <linux/magic.h>
  36 #include <linux/migrate.h>
  37 #include <linux/uio.h>
  38
  39 #include <asm/uaccess.h>
  40
  41 static const struct super_operations hugetlbfs_ops;
  42 static const struct address_space_operations hugetlbfs_aops;
  43 const struct file_operations hugetlbfs_file_operations;
  44 static const struct inode_operations hugetlbfs_dir_inode_operations;
  45 static const struct inode_operations hugetlbfs_inode_operations;
  46
  47 struct hugetlbfs_config {
  48         kuid_t   uid;
  49         kgid_t   gid;
  50         umode_t mode;
  51         long    max_hpages;
  52         long    nr_inodes;
  53         struct hstate *hstate;
  54         long    min_hpages;
  55 };
  56
  57 struct hugetlbfs_inode_info {
  58         struct shared_policy policy;
  59         struct inode vfs_inode;
  60 };
  61
  62 static inline struct hugetlbfs_inode_info *HUGETLBFS_I(struct inode *inode)
  63 {
  64         return container_of(inode, struct hugetlbfs_inode_info, vfs_inode);
  65 }
  66
  67 int sysctl_hugetlb_shm_group;
  68
  69 enum {
  70         Opt_size, Opt_nr_inodes,
  71         Opt_mode, Opt_uid, Opt_gid,
  72         Opt_pagesize, Opt_min_size,
  73         Opt_err,
  74 };
  75
  76 static const match_table_t tokens = {
  77         {Opt_size,      "size=%s"},
  78         {Opt_nr_inodes, "nr_inodes=%s"},
  79         {Opt_mode,      "mode=%o"},
  80         {Opt_uid,       "uid=%u"},
  81         {Opt_gid,       "gid=%u"},
  82         {Opt_pagesize,  "pagesize=%s"},
  83         {Opt_min_size,  "min_size=%s"},
  84         {Opt_err,       NULL},
  85 };
  86
  87 static void huge_pagevec_release(struct pagevec *pvec)
  88 {
  89         int i;
  90
  91         for (i = 0; i < pagevec_count(pvec); ++i)
  92                 put_page(pvec->pages[i]);
  93
  94         pagevec_reinit(pvec);
  95 }
  96
  97 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
  98 {
  99         struct inode *inode = file_inode(file);
 100         loff_t len, vma_len;
 101         int ret;
 102         struct hstate *h = hstate_file(file);
 103
 104         /*
 105          * vma address alignment (but not the pgoff alignment) has
 106          * already been checked by prepare_hugepage_range.  If you add
 107          * any error returns here, do so after setting VM_HUGETLB, so
 108          * is_vm_hugetlb_page tests below unmap_region go the right
 109          * way when do_mmap_pgoff unwinds (may be important on powerpc
 110          * and ia64).
 111          */
 112         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 113         vma->vm_ops = &hugetlb_vm_ops;
 114
 115         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 116                 return -EINVAL;
 117
 118         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 119
 120         mutex_lock(&inode->i_mutex);
 121         file_accessed(file);
 122
 123         ret = -ENOMEM;
 124         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 125
 126         if (hugetlb_reserve_pages(inode,
 127                                 vma->vm_pgoff >> huge_page_order(h),
 128                                 len >> huge_page_shift(h), vma,
 129                                 vma->vm_flags))
 130                 goto out;
 131
 132         ret = 0;
 133         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 134                 inode->i_size = len;
 135 out:
 136         mutex_unlock(&inode->i_mutex);
 137
 138         return ret;
 139 }
 140
 141 /*
 142  * Called under down_write(mmap_sem).
 143  */
 144
 145 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 146 static unsigned long
 147 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 148                 unsigned long len, unsigned long pgoff, unsigned long flags)
 149 {
 150         struct mm_struct *mm = current->mm;
 151         struct vm_area_struct *vma;
 152         struct hstate *h = hstate_file(file);
 153         struct vm_unmapped_area_info info;
 154
 155         if (len & ~huge_page_mask(h))
 156                 return -EINVAL;
 157         if (len > TASK_SIZE)
 158                 return -ENOMEM;
 159
 160         if (flags & MAP_FIXED) {
 161                 if (prepare_hugepage_range(file, addr, len))
 162                         return -EINVAL;
 163                 return addr;
 164         }
 165
 166         if (addr) {
 167                 addr = ALIGN(addr, huge_page_size(h));
 168                 vma = find_vma(mm, addr);
 169                 if (TASK_SIZE - len >= addr &&
 170                     (!vma || addr + len <= vma->vm_start))
 171                         return addr;
 172         }
 173
 174         info.flags = 0;
 175         info.length = len;
 176         info.low_limit = TASK_UNMAPPED_BASE;
 177         info.high_limit = TASK_SIZE;
 178         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 179         info.align_offset = 0;
 180         return vm_unmapped_area(&info);
 181 }
 182 #endif
 183
 184 static size_t
 185 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 186                         struct iov_iter *to, unsigned long size)
 187 {
 188         size_t copied = 0;
 189         int i, chunksize;
 190
 191         /* Find which 4k chunk and offset with in that chunk */
 192         i = offset >> PAGE_CACHE_SHIFT;
 193         offset = offset & ~PAGE_CACHE_MASK;
 194
 195         while (size) {
 196                 size_t n;
 197                 chunksize = PAGE_CACHE_SIZE;
 198                 if (offset)
 199                         chunksize -= offset;
 200                 if (chunksize > size)
 201                         chunksize = size;
 202                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 203                 copied += n;
 204                 if (n != chunksize)
 205                         return copied;
 206                 offset = 0;
 207                 size -= chunksize;
 208                 i++;
 209         }
 210         return copied;
 211 }
 212
 213 /*
 214  * Support for read() - Find the page attached to f_mapping and copy out the
 215  * data. Its *very* similar to do_generic_mapping_read(), we can't use that
 216  * since it has PAGE_CACHE_SIZE assumptions.
 217  */
 218 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 219 {
 220         struct file *file = iocb->ki_filp;
 221         struct hstate *h = hstate_file(file);
 222         struct address_space *mapping = file->f_mapping;
 223         struct inode *inode = mapping->host;
 224         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 225         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 226         unsigned long end_index;
 227         loff_t isize;
 228         ssize_t retval = 0;
 229
 230         while (iov_iter_count(to)) {
 231                 struct page *page;
 232                 size_t nr, copied;
 233
 234                 /* nr is the maximum number of bytes to copy from this page */
 235                 nr = huge_page_size(h);
 236                 isize = i_size_read(inode);
 237                 if (!isize)
 238                         break;
 239                 end_index = (isize - 1) >> huge_page_shift(h);
 240                 if (index > end_index)
 241                         break;
 242                 if (index == end_index) {
 243                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 244                         if (nr <= offset)
 245                                 break;
 246                 }
 247                 nr = nr - offset;
 248
 249                 /* Find the page */
 250                 page = find_lock_page(mapping, index);
 251                 if (unlikely(page == NULL)) {
 252                         /*
 253                          * We have a HOLE, zero out the user-buffer for the
 254                          * length of the hole or request.
 255                          */
 256                         copied = iov_iter_zero(nr, to);
 257                 } else {
 258                         unlock_page(page);
 259
 260                         /*
 261                          * We have the page, copy it to user space buffer.
 262                          */
 263                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 264                         page_cache_release(page);
 265                 }
 266                 offset += copied;
 267                 retval += copied;
 268                 if (copied != nr && iov_iter_count(to)) {
 269                         if (!retval)
 270                                 retval = -EFAULT;
 271                         break;
 272                 }
 273                 index += offset >> huge_page_shift(h);
 274                 offset &= ~huge_page_mask(h);
 275         }
 276         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 277         return retval;
 278 }
 279
 280 static int hugetlbfs_write_begin(struct file *file,
 281                         struct address_space *mapping,
 282                         loff_t pos, unsigned len, unsigned flags,
 283                         struct page **pagep, void **fsdata)
 284 {
 285         return -EINVAL;
 286 }
 287
 288 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 289                         loff_t pos, unsigned len, unsigned copied,
 290                         struct page *page, void *fsdata)
 291 {
 292         BUG();
 293         return -EINVAL;
 294 }
 295
 296 static void truncate_huge_page(struct page *page)
 297 {
 298         ClearPageDirty(page);
 299         ClearPageUptodate(page);
 300         delete_from_page_cache(page);
 301 }
 302
 303 static void truncate_hugepages(struct inode *inode, loff_t lstart)
 304 {
 305         struct hstate *h = hstate_inode(inode);
 306         struct address_space *mapping = &inode->i_data;
 307         const pgoff_t start = lstart >> huge_page_shift(h);
 308         struct pagevec pvec;
 309         pgoff_t next;
 310         int i, freed = 0;
 311
 312         pagevec_init(&pvec, 0);
 313         next = start;
 314         while (1) {
 315                 if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
 316                         if (next == start)
 317                                 break;
 318                         next = start;
 319                         continue;
 320                 }
 321
 322                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 323                         struct page *page = pvec.pages[i];
 324
 325                         lock_page(page);
 326                         if (page->index > next)
 327                                 next = page->index;
 328                         ++next;
 329                         truncate_huge_page(page);
 330                         unlock_page(page);
 331                         freed++;
 332                 }
 333                 huge_pagevec_release(&pvec);
 334         }
 335         BUG_ON(!lstart && mapping->nrpages);
 336         hugetlb_unreserve_pages(inode, start, freed);
 337 }
 338
 339 static void hugetlbfs_evict_inode(struct inode *inode)
 340 {
 341         struct resv_map *resv_map;
 342
 343         truncate_hugepages(inode, 0);
 344         resv_map = (struct resv_map *)inode->i_mapping->private_data;
 345         /* root inode doesn't have the resv_map, so we should check it */
 346         if (resv_map)
 347                 resv_map_release(&resv_map->refs);
 348         clear_inode(inode);
 349 }
 350
 351 static inline void
 352 hugetlb_vmdelete_list(struct rb_root *root, pgoff_t start, pgoff_t end)
 353 {
 354         struct vm_area_struct *vma;
 355
 356         /*
 357          * end == 0 indicates that the entire range after
 358          * start should be unmapped.
 359          */
 360         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 361                 unsigned long v_offset;
 362
 363                 /*
 364                  * Can the expression below overflow on 32-bit arches?
 365                  * No, because the interval tree returns us only those vmas
 366                  * which overlap the truncated area starting at pgoff,
 367                  * and no vma on a 32-bit arch can span beyond the 4GB.
 368                  */
 369                 if (vma->vm_pgoff < start)
 370                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 371                 else
 372                         v_offset = 0;
 373
 374                 if (end) {
 375                         end = ((end - start) << PAGE_SHIFT) +
 376                                vma->vm_start + v_offset;
 377                         if (end > vma->vm_end)
 378                                 end = vma->vm_end;
 379                 } else
 380                         end = vma->vm_end;
 381
 382                 unmap_hugepage_range(vma, vma->vm_start + v_offset, end, NULL);
 383         }
 384 }
 385
 386 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 387 {
 388         pgoff_t pgoff;
 389         struct address_space *mapping = inode->i_mapping;
 390         struct hstate *h = hstate_inode(inode);
 391
 392         BUG_ON(offset & ~huge_page_mask(h));
 393         pgoff = offset >> PAGE_SHIFT;
 394
 395         i_size_write(inode, offset);
 396         i_mmap_lock_write(mapping);
 397         if (!RB_EMPTY_ROOT(&mapping->i_mmap))
 398                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 399         i_mmap_unlock_write(mapping);
 400         truncate_hugepages(inode, offset);
 401         return 0;
 402 }
 403
 404 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 405 {
 406         struct inode *inode = d_inode(dentry);
 407         struct hstate *h = hstate_inode(inode);
 408         int error;
 409         unsigned int ia_valid = attr->ia_valid;
 410
 411         BUG_ON(!inode);
 412
 413         error = inode_change_ok(inode, attr);
 414         if (error)
 415                 return error;
 416
 417         if (ia_valid & ATTR_SIZE) {
 418                 error = -EINVAL;
 419                 if (attr->ia_size & ~huge_page_mask(h))
 420                         return -EINVAL;
 421                 error = hugetlb_vmtruncate(inode, attr->ia_size);
 422                 if (error)
 423                         return error;
 424         }
 425
 426         setattr_copy(inode, attr);
 427         mark_inode_dirty(inode);
 428         return 0;
 429 }
 430
 431 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 432                                         struct hugetlbfs_config *config)
 433 {
 434         struct inode *inode;
 435
 436         inode = new_inode(sb);
 437         if (inode) {
 438                 struct hugetlbfs_inode_info *info;
 439                 inode->i_ino = get_next_ino();
 440                 inode->i_mode = S_IFDIR | config->mode;
 441                 inode->i_uid = config->uid;
 442                 inode->i_gid = config->gid;
 443                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 444                 info = HUGETLBFS_I(inode);
 445                 mpol_shared_policy_init(&info->policy, NULL);
 446                 inode->i_op = &hugetlbfs_dir_inode_operations;
 447                 inode->i_fop = &simple_dir_operations;
 448                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 449                 inc_nlink(inode);
 450                 lockdep_annotate_inode_mutex_key(inode);
 451         }
 452         return inode;
 453 }
 454
 455 /*
 456  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 457  * be taken from reclaim -- unlike regular filesystems. This needs an
 458  * annotation because huge_pmd_share() does an allocation under
 459  * i_mmap_rwsem.
 460  */
 461 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 462
 463 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 464                                         struct inode *dir,
 465                                         umode_t mode, dev_t dev)
 466 {
 467         struct inode *inode;
 468         struct resv_map *resv_map;
 469
 470         resv_map = resv_map_alloc();
 471         if (!resv_map)
 472                 return NULL;
 473
 474         inode = new_inode(sb);
 475         if (inode) {
 476                 struct hugetlbfs_inode_info *info;
 477                 inode->i_ino = get_next_ino();
 478                 inode_init_owner(inode, dir, mode);
 479                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 480                                 &hugetlbfs_i_mmap_rwsem_key);
 481                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 482                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 483                 inode->i_mapping->private_data = resv_map;
 484                 info = HUGETLBFS_I(inode);
 485                 /*
 486                  * The policy is initialized here even if we are creating a
 487                  * private inode because initialization simply creates an
 488                  * an empty rb tree and calls spin_lock_init(), later when we
 489                  * call mpol_free_shared_policy() it will just return because
 490                  * the rb tree will still be empty.
 491                  */
 492                 mpol_shared_policy_init(&info->policy, NULL);
 493                 switch (mode & S_IFMT) {
 494                 default:
 495                         init_special_inode(inode, mode, dev);
 496                         break;
 497                 case S_IFREG:
 498                         inode->i_op = &hugetlbfs_inode_operations;
 499                         inode->i_fop = &hugetlbfs_file_operations;
 500                         break;
 501                 case S_IFDIR:
 502                         inode->i_op = &hugetlbfs_dir_inode_operations;
 503                         inode->i_fop = &simple_dir_operations;
 504
 505                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 506                         inc_nlink(inode);
 507                         break;
 508                 case S_IFLNK:
 509                         inode->i_op = &page_symlink_inode_operations;
 510                         break;
 511                 }
 512                 lockdep_annotate_inode_mutex_key(inode);
 513         } else
 514                 kref_put(&resv_map->refs, resv_map_release);
 515
 516         return inode;
 517 }
 518
 519 /*
 520  * File creation. Allocate an inode, and we're done..
 521  */
 522 static int hugetlbfs_mknod(struct inode *dir,
 523                         struct dentry *dentry, umode_t mode, dev_t dev)
 524 {
 525         struct inode *inode;
 526         int error = -ENOSPC;
 527
 528         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 529         if (inode) {
 530                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 531                 d_instantiate(dentry, inode);
 532                 dget(dentry);   /* Extra count - pin the dentry in core */
 533                 error = 0;
 534         }
 535         return error;
 536 }
 537
 538 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 539 {
 540         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 541         if (!retval)
 542                 inc_nlink(dir);
 543         return retval;
 544 }
 545
 546 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 547 {
 548         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 549 }
 550
 551 static int hugetlbfs_symlink(struct inode *dir,
 552                         struct dentry *dentry, const char *symname)
 553 {
 554         struct inode *inode;
 555         int error = -ENOSPC;
 556
 557         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 558         if (inode) {
 559                 int l = strlen(symname)+1;
 560                 error = page_symlink(inode, symname, l);
 561                 if (!error) {
 562                         d_instantiate(dentry, inode);
 563                         dget(dentry);
 564                 } else
 565                         iput(inode);
 566         }
 567         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 568
 569         return error;
 570 }
 571
 572 /*
 573  * mark the head page dirty
 574  */
 575 static int hugetlbfs_set_page_dirty(struct page *page)
 576 {
 577         struct page *head = compound_head(page);
 578
 579         SetPageDirty(head);
 580         return 0;
 581 }
 582
 583 static int hugetlbfs_migrate_page(struct address_space *mapping,
 584                                 struct page *newpage, struct page *page,
 585                                 enum migrate_mode mode)
 586 {
 587         int rc;
 588
 589         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 590         if (rc != MIGRATEPAGE_SUCCESS)
 591                 return rc;
 592         migrate_page_copy(newpage, page);
 593
 594         return MIGRATEPAGE_SUCCESS;
 595 }
 596
 597 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 598 {
 599         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 600         struct hstate *h = hstate_inode(d_inode(dentry));
 601
 602         buf->f_type = HUGETLBFS_MAGIC;
 603         buf->f_bsize = huge_page_size(h);
 604         if (sbinfo) {
 605                 spin_lock(&sbinfo->stat_lock);
 606                 /* If no limits set, just report 0 for max/free/used
 607                  * blocks, like simple_statfs() */
 608                 if (sbinfo->spool) {
 609                         long free_pages;
 610
 611                         spin_lock(&sbinfo->spool->lock);
 612                         buf->f_blocks = sbinfo->spool->max_hpages;
 613                         free_pages = sbinfo->spool->max_hpages
 614                                 - sbinfo->spool->used_hpages;
 615                         buf->f_bavail = buf->f_bfree = free_pages;
 616                         spin_unlock(&sbinfo->spool->lock);
 617                         buf->f_files = sbinfo->max_inodes;
 618                         buf->f_ffree = sbinfo->free_inodes;
 619                 }
 620                 spin_unlock(&sbinfo->stat_lock);
 621         }
 622         buf->f_namelen = NAME_MAX;
 623         return 0;
 624 }
 625
 626 static void hugetlbfs_put_super(struct super_block *sb)
 627 {
 628         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
 629
 630         if (sbi) {
 631                 sb->s_fs_info = NULL;
 632
 633                 if (sbi->spool)
 634                         hugepage_put_subpool(sbi->spool);
 635
 636                 kfree(sbi);
 637         }
 638 }
 639
 640 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 641 {
 642         if (sbinfo->free_inodes >= 0) {
 643                 spin_lock(&sbinfo->stat_lock);
 644                 if (unlikely(!sbinfo->free_inodes)) {
 645                         spin_unlock(&sbinfo->stat_lock);
 646                         return 0;
 647                 }
 648                 sbinfo->free_inodes--;
 649                 spin_unlock(&sbinfo->stat_lock);
 650         }
 651
 652         return 1;
 653 }
 654
 655 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 656 {
 657         if (sbinfo->free_inodes >= 0) {
 658                 spin_lock(&sbinfo->stat_lock);
 659                 sbinfo->free_inodes++;
 660                 spin_unlock(&sbinfo->stat_lock);
 661         }
 662 }
 663
 664
 665 static struct kmem_cache *hugetlbfs_inode_cachep;
 666
 667 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
 668 {
 669         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
 670         struct hugetlbfs_inode_info *p;
 671
 672         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
 673                 return NULL;
 674         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
 675         if (unlikely(!p)) {
 676                 hugetlbfs_inc_free_inodes(sbinfo);
 677                 return NULL;
 678         }
 679         return &p->vfs_inode;
 680 }
 681
 682 static void hugetlbfs_i_callback(struct rcu_head *head)
 683 {
 684         struct inode *inode = container_of(head, struct inode, i_rcu);
 685         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
 686 }
 687
 688 static void hugetlbfs_destroy_inode(struct inode *inode)
 689 {
 690         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
 691         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
 692         call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
 693 }
 694
 695 static const struct address_space_operations hugetlbfs_aops = {
 696         .write_begin    = hugetlbfs_write_begin,
 697         .write_end      = hugetlbfs_write_end,
 698         .set_page_dirty = hugetlbfs_set_page_dirty,
 699         .migratepage    = hugetlbfs_migrate_page,
 700 };
 701
 702
 703 static void init_once(void *foo)
 704 {
 705         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
 706
 707         inode_init_once(&ei->vfs_inode);
 708 }
 709
 710 const struct file_operations hugetlbfs_file_operations = {
 711         .read_iter              = hugetlbfs_read_iter,
 712         .mmap                   = hugetlbfs_file_mmap,
 713         .fsync                  = noop_fsync,
 714         .get_unmapped_area      = hugetlb_get_unmapped_area,
 715         .llseek         = default_llseek,
 716 };
 717
 718 static const struct inode_operations hugetlbfs_dir_inode_operations = {
 719         .create         = hugetlbfs_create,
 720         .lookup         = simple_lookup,
 721         .link           = simple_link,
 722         .unlink         = simple_unlink,
 723         .symlink        = hugetlbfs_symlink,
 724         .mkdir          = hugetlbfs_mkdir,
 725         .rmdir          = simple_rmdir,
 726         .mknod          = hugetlbfs_mknod,
 727         .rename         = simple_rename,
 728         .setattr        = hugetlbfs_setattr,
 729 };
 730
 731 static const struct inode_operations hugetlbfs_inode_operations = {
 732         .setattr        = hugetlbfs_setattr,
 733 };
 734
 735 static const struct super_operations hugetlbfs_ops = {
 736         .alloc_inode    = hugetlbfs_alloc_inode,
 737         .destroy_inode  = hugetlbfs_destroy_inode,
 738         .evict_inode    = hugetlbfs_evict_inode,
 739         .statfs         = hugetlbfs_statfs,
 740         .put_super      = hugetlbfs_put_super,
 741         .show_options   = generic_show_options,
 742 };
 743
 744 enum { NO_SIZE, SIZE_STD, SIZE_PERCENT };
 745
 746 /*
 747  * Convert size option passed from command line to number of huge pages
 748  * in the pool specified by hstate.  Size option could be in bytes
 749  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
 750  */
 751 static long long
 752 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
 753                                                                 int val_type)
 754 {
 755         if (val_type == NO_SIZE)
 756                 return -1;
 757
 758         if (val_type == SIZE_PERCENT) {
 759                 size_opt <<= huge_page_shift(h);
 760                 size_opt *= h->max_huge_pages;
 761                 do_div(size_opt, 100);
 762         }
 763
 764         size_opt >>= huge_page_shift(h);
 765         return size_opt;
 766 }
 767
 768 static int
 769 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
 770 {
 771         char *p, *rest;
 772         substring_t args[MAX_OPT_ARGS];
 773         int option;
 774         unsigned long long max_size_opt = 0, min_size_opt = 0;
 775         int max_val_type = NO_SIZE, min_val_type = NO_SIZE;
 776
 777         if (!options)
 778                 return 0;
 779
 780         while ((p = strsep(&options, ",")) != NULL) {
 781                 int token;
 782                 if (!*p)
 783                         continue;
 784
 785                 token = match_token(p, tokens, args);
 786                 switch (token) {
 787                 case Opt_uid:
 788                         if (match_int(&args[0], &option))
 789                                 goto bad_val;
 790                         pconfig->uid = make_kuid(current_user_ns(), option);
 791                         if (!uid_valid(pconfig->uid))
 792                                 goto bad_val;
 793                         break;
 794
 795                 case Opt_gid:
 796                         if (match_int(&args[0], &option))
 797                                 goto bad_val;
 798                         pconfig->gid = make_kgid(current_user_ns(), option);
 799                         if (!gid_valid(pconfig->gid))
 800                                 goto bad_val;
 801                         break;
 802
 803                 case Opt_mode:
 804                         if (match_octal(&args[0], &option))
 805                                 goto bad_val;
 806                         pconfig->mode = option & 01777U;
 807                         break;
 808
 809                 case Opt_size: {
 810                         /* memparse() will accept a K/M/G without a digit */
 811                         if (!isdigit(*args[0].from))
 812                                 goto bad_val;
 813                         max_size_opt = memparse(args[0].from, &rest);
 814                         max_val_type = SIZE_STD;
 815                         if (*rest == '%')
 816                                 max_val_type = SIZE_PERCENT;
 817                         break;
 818                 }
 819
 820                 case Opt_nr_inodes:
 821                         /* memparse() will accept a K/M/G without a digit */
 822                         if (!isdigit(*args[0].from))
 823                                 goto bad_val;
 824                         pconfig->nr_inodes = memparse(args[0].from, &rest);
 825                         break;
 826
 827                 case Opt_pagesize: {
 828                         unsigned long ps;
 829                         ps = memparse(args[0].from, &rest);
 830                         pconfig->hstate = size_to_hstate(ps);
 831                         if (!pconfig->hstate) {
 832                                 pr_err("Unsupported page size %lu MB\n",
 833                                         ps >> 20);
 834                                 return -EINVAL;
 835                         }
 836                         break;
 837                 }
 838
 839                 case Opt_min_size: {
 840                         /* memparse() will accept a K/M/G without a digit */
 841                         if (!isdigit(*args[0].from))
 842                                 goto bad_val;
 843                         min_size_opt = memparse(args[0].from, &rest);
 844                         min_val_type = SIZE_STD;
 845                         if (*rest == '%')
 846                                 min_val_type = SIZE_PERCENT;
 847                         break;
 848                 }
 849
 850                 default:
 851                         pr_err("Bad mount option: \"%s\"\n", p);
 852                         return -EINVAL;
 853                         break;
 854                 }
 855         }
 856
 857         /*
 858          * Use huge page pool size (in hstate) to convert the size
 859          * options to number of huge pages.  If NO_SIZE, -1 is returned.
 860          */
 861         pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
 862                                                 max_size_opt, max_val_type);
 863         pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
 864                                                 min_size_opt, min_val_type);
 865
 866         /*
 867          * If max_size was specified, then min_size must be smaller
 868          */
 869         if (max_val_type > NO_SIZE &&
 870             pconfig->min_hpages > pconfig->max_hpages) {
 871                 pr_err("minimum size can not be greater than maximum size\n");
 872                 return -EINVAL;
 873         }
 874
 875         return 0;
 876
 877 bad_val:
 878         pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
 879         return -EINVAL;
 880 }
 881
 882 static int
 883 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
 884 {
 885         int ret;
 886         struct hugetlbfs_config config;
 887         struct hugetlbfs_sb_info *sbinfo;
 888
 889         save_mount_options(sb, data);
 890
 891         config.max_hpages = -1; /* No limit on size by default */
 892         config.nr_inodes = -1; /* No limit on number of inodes by default */
 893         config.uid = current_fsuid();
 894         config.gid = current_fsgid();
 895         config.mode = 0755;
 896         config.hstate = &default_hstate;
 897         config.min_hpages = -1; /* No default minimum size */
 898         ret = hugetlbfs_parse_options(data, &config);
 899         if (ret)
 900                 return ret;
 901
 902         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
 903         if (!sbinfo)
 904                 return -ENOMEM;
 905         sb->s_fs_info = sbinfo;
 906         sbinfo->hstate = config.hstate;
 907         spin_lock_init(&sbinfo->stat_lock);
 908         sbinfo->max_inodes = config.nr_inodes;
 909         sbinfo->free_inodes = config.nr_inodes;
 910         sbinfo->spool = NULL;
 911         /*
 912          * Allocate and initialize subpool if maximum or minimum size is
 913          * specified.  Any needed reservations (for minimim size) are taken
 914          * taken when the subpool is created.
 915          */
 916         if (config.max_hpages != -1 || config.min_hpages != -1) {
 917                 sbinfo->spool = hugepage_new_subpool(config.hstate,
 918                                                         config.max_hpages,
 919                                                         config.min_hpages);
 920                 if (!sbinfo->spool)
 921                         goto out_free;
 922         }
 923         sb->s_maxbytes = MAX_LFS_FILESIZE;
 924         sb->s_blocksize = huge_page_size(config.hstate);
 925         sb->s_blocksize_bits = huge_page_shift(config.hstate);
 926         sb->s_magic = HUGETLBFS_MAGIC;
 927         sb->s_op = &hugetlbfs_ops;
 928         sb->s_time_gran = 1;
 929         sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
 930         if (!sb->s_root)
 931                 goto out_free;
 932         return 0;
 933 out_free:
 934         kfree(sbinfo->spool);
 935         kfree(sbinfo);
 936         return -ENOMEM;
 937 }
 938
 939 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
 940         int flags, const char *dev_name, void *data)
 941 {
 942         return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
 943 }
 944
 945 static struct file_system_type hugetlbfs_fs_type = {
 946         .name           = "hugetlbfs",
 947         .mount          = hugetlbfs_mount,
 948         .kill_sb        = kill_litter_super,
 949 };
 950 MODULE_ALIAS_FS("hugetlbfs");
 951
 952 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
 953
 954 static int can_do_hugetlb_shm(void)
 955 {
 956         kgid_t shm_group;
 957         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
 958         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
 959 }
 960
 961 static int get_hstate_idx(int page_size_log)
 962 {
 963         struct hstate *h = hstate_sizelog(page_size_log);
 964
 965         if (!h)
 966                 return -1;
 967         return h - hstates;
 968 }
 969
 970 static const struct dentry_operations anon_ops = {
 971         .d_dname = simple_dname
 972 };
 973
 974 /*
 975  * Note that size should be aligned to proper hugepage size in caller side,
 976  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
 977  */
 978 struct file *hugetlb_file_setup(const char *name, size_t size,
 979                                 vm_flags_t acctflag, struct user_struct **user,
 980                                 int creat_flags, int page_size_log)
 981 {
 982         struct file *file = ERR_PTR(-ENOMEM);
 983         struct inode *inode;
 984         struct path path;
 985         struct super_block *sb;
 986         struct qstr quick_string;
 987         int hstate_idx;
 988
 989         hstate_idx = get_hstate_idx(page_size_log);
 990         if (hstate_idx < 0)
 991                 return ERR_PTR(-ENODEV);
 992
 993         *user = NULL;
 994         if (!hugetlbfs_vfsmount[hstate_idx])
 995                 return ERR_PTR(-ENOENT);
 996
 997         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
 998                 *user = current_user();
 999                 if (user_shm_lock(size, *user)) {
1000                         task_lock(current);
1001                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1002                                 current->comm, current->pid);
1003                         task_unlock(current);
1004                 } else {
1005                         *user = NULL;
1006                         return ERR_PTR(-EPERM);
1007                 }
1008         }
1009
1010         sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
1011         quick_string.name = name;
1012         quick_string.len = strlen(quick_string.name);
1013         quick_string.hash = 0;
1014         path.dentry = d_alloc_pseudo(sb, &quick_string);
1015         if (!path.dentry)
1016                 goto out_shm_unlock;
1017
1018         d_set_d_op(path.dentry, &anon_ops);
1019         path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
1020         file = ERR_PTR(-ENOSPC);
1021         inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
1022         if (!inode)
1023                 goto out_dentry;
1024         if (creat_flags == HUGETLB_SHMFS_INODE)
1025                 inode->i_flags |= S_PRIVATE;
1026
1027         file = ERR_PTR(-ENOMEM);
1028         if (hugetlb_reserve_pages(inode, 0,
1029                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1030                         acctflag))
1031                 goto out_inode;
1032
1033         d_instantiate(path.dentry, inode);
1034         inode->i_size = size;
1035         clear_nlink(inode);
1036
1037         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1038                         &hugetlbfs_file_operations);
1039         if (IS_ERR(file))
1040                 goto out_dentry; /* inode is already attached */
1041
1042         return file;
1043
1044 out_inode:
1045         iput(inode);
1046 out_dentry:
1047         path_put(&path);
1048 out_shm_unlock:
1049         if (*user) {
1050                 user_shm_unlock(size, *user);
1051                 *user = NULL;
1052         }
1053         return file;
1054 }
1055
1056 static int __init init_hugetlbfs_fs(void)
1057 {
1058         struct hstate *h;
1059         int error;
1060         int i;
1061
1062         if (!hugepages_supported()) {
1063                 pr_info("disabling because there are no supported hugepage sizes\n");
1064                 return -ENOTSUPP;
1065         }
1066
1067         error = -ENOMEM;
1068         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1069                                         sizeof(struct hugetlbfs_inode_info),
1070                                         0, 0, init_once);
1071         if (hugetlbfs_inode_cachep == NULL)
1072                 goto out2;
1073
1074         error = register_filesystem(&hugetlbfs_fs_type);
1075         if (error)
1076                 goto out;
1077
1078         i = 0;
1079         for_each_hstate(h) {
1080                 char buf[50];
1081                 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1082
1083                 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1084                 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1085                                                         buf);
1086
1087                 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1088                         pr_err("Cannot mount internal hugetlbfs for "
1089                                 "page size %uK", ps_kb);
1090                         error = PTR_ERR(hugetlbfs_vfsmount[i]);
1091                         hugetlbfs_vfsmount[i] = NULL;
1092                 }
1093                 i++;
1094         }
1095         /* Non default hstates are optional */
1096         if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1097                 return 0;
1098
1099  out:
1100         kmem_cache_destroy(hugetlbfs_inode_cachep);
1101  out2:
1102         return error;
1103 }
1104
1105 static void __exit exit_hugetlbfs_fs(void)
1106 {
1107         struct hstate *h;
1108         int i;
1109
1110
1111         /*
1112          * Make sure all delayed rcu free inodes are flushed before we
1113          * destroy cache.
1114          */
1115         rcu_barrier();
1116         kmem_cache_destroy(hugetlbfs_inode_cachep);
1117         i = 0;
1118         for_each_hstate(h)
1119                 kern_unmount(hugetlbfs_vfsmount[i++]);
1120         unregister_filesystem(&hugetlbfs_fs_type);
1121 }
1122
1123 module_init(init_hugetlbfs_fs)
1124 module_exit(exit_hugetlbfs_fs)
1125
1126 MODULE_LICENSE("GPL");