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