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