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