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    max_hpages;
  51         long    nr_inodes;
  52         struct hstate *hstate;
  53         long    min_hpages;
  54 };
  55
  56 struct hugetlbfs_inode_info {
  57         struct shared_policy policy;
  58         struct inode vfs_inode;
  59 };
  60
  61 static inline struct hugetlbfs_inode_info *HUGETLBFS_I(struct inode *inode)
  62 {
  63         return container_of(inode, struct hugetlbfs_inode_info, vfs_inode);
  64 }
  65
  66 int sysctl_hugetlb_shm_group;
  67
  68 enum {
  69         Opt_size, Opt_nr_inodes,
  70         Opt_mode, Opt_uid, Opt_gid,
  71         Opt_pagesize, Opt_min_size,
  72         Opt_err,
  73 };
  74
  75 static const match_table_t tokens = {
  76         {Opt_size,      "size=%s"},
  77         {Opt_nr_inodes, "nr_inodes=%s"},
  78         {Opt_mode,      "mode=%o"},
  79         {Opt_uid,       "uid=%u"},
  80         {Opt_gid,       "gid=%u"},
  81         {Opt_pagesize,  "pagesize=%s"},
  82         {Opt_min_size,  "min_size=%s"},
  83         {Opt_err,       NULL},
  84 };
  85
  86 static void huge_pagevec_release(struct pagevec *pvec)
  87 {
  88         int i;
  89
  90         for (i = 0; i < pagevec_count(pvec); ++i)
  91                 put_page(pvec->pages[i]);
  92
  93         pagevec_reinit(pvec);
  94 }
  95
  96 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
  97 {
  98         struct inode *inode = file_inode(file);
  99         loff_t len, vma_len;
 100         int ret;
 101         struct hstate *h = hstate_file(file);
 102
 103         /*
 104          * vma address alignment (but not the pgoff alignment) has
 105          * already been checked by prepare_hugepage_range.  If you add
 106          * any error returns here, do so after setting VM_HUGETLB, so
 107          * is_vm_hugetlb_page tests below unmap_region go the right
 108          * way when do_mmap_pgoff unwinds (may be important on powerpc
 109          * and ia64).
 110          */
 111         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 112         vma->vm_ops = &hugetlb_vm_ops;
 113
 114         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 115                 return -EINVAL;
 116
 117         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 118
 119         mutex_lock(&inode->i_mutex);
 120         file_accessed(file);
 121
 122         ret = -ENOMEM;
 123         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 124
 125         if (hugetlb_reserve_pages(inode,
 126                                 vma->vm_pgoff >> huge_page_order(h),
 127                                 len >> huge_page_shift(h), vma,
 128                                 vma->vm_flags))
 129                 goto out;
 130
 131         ret = 0;
 132         hugetlb_prefault_arch_hook(vma->vm_mm);
 133         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 134                 inode->i_size = len;
 135 out:
 136         mutex_unlock(&inode->i_mutex);
 137
 138         return ret;
 139 }
 140
 141 /*
 142  * Called under down_write(mmap_sem).
 143  */
 144
 145 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 146 static unsigned long
 147 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 148                 unsigned long len, unsigned long pgoff, unsigned long flags)
 149 {
 150         struct mm_struct *mm = current->mm;
 151         struct vm_area_struct *vma;
 152         struct hstate *h = hstate_file(file);
 153         struct vm_unmapped_area_info info;
 154
 155         if (len & ~huge_page_mask(h))
 156                 return -EINVAL;
 157         if (len > TASK_SIZE)
 158                 return -ENOMEM;
 159
 160         if (flags & MAP_FIXED) {
 161                 if (prepare_hugepage_range(file, addr, len))
 162                         return -EINVAL;
 163                 return addr;
 164         }
 165
 166         if (addr) {
 167                 addr = ALIGN(addr, huge_page_size(h));
 168                 vma = find_vma(mm, addr);
 169                 if (TASK_SIZE - len >= addr &&
 170                     (!vma || addr + len <= vma->vm_start))
 171                         return addr;
 172         }
 173
 174         info.flags = 0;
 175         info.length = len;
 176         info.low_limit = TASK_UNMAPPED_BASE;
 177         info.high_limit = TASK_SIZE;
 178         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 179         info.align_offset = 0;
 180         return vm_unmapped_area(&info);
 181 }
 182 #endif
 183
 184 static int
 185 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 186                         char __user *buf, unsigned long count,
 187                         unsigned long size)
 188 {
 189         char *kaddr;
 190         unsigned long left, copied = 0;
 191         int i, chunksize;
 192
 193         if (size > count)
 194                 size = count;
 195
 196         /* Find which 4k chunk and offset with in that chunk */
 197         i = offset >> PAGE_CACHE_SHIFT;
 198         offset = offset & ~PAGE_CACHE_MASK;
 199
 200         while (size) {
 201                 chunksize = PAGE_CACHE_SIZE;
 202                 if (offset)
 203                         chunksize -= offset;
 204                 if (chunksize > size)
 205                         chunksize = size;
 206                 kaddr = kmap(&page[i]);
 207                 left = __copy_to_user(buf, kaddr + offset, chunksize);
 208                 kunmap(&page[i]);
 209                 if (left) {
 210                         copied += (chunksize - left);
 211                         break;
 212                 }
 213                 offset = 0;
 214                 size -= chunksize;
 215                 buf += chunksize;
 216                 copied += chunksize;
 217                 i++;
 218         }
 219         return copied ? copied : -EFAULT;
 220 }
 221
 222 /*
 223  * Support for read() - Find the page attached to f_mapping and copy out the
 224  * data. Its *very* similar to do_generic_mapping_read(), we can't use that
 225  * since it has PAGE_CACHE_SIZE assumptions.
 226  */
 227 static ssize_t hugetlbfs_read(struct file *filp, char __user *buf,
 228                               size_t len, loff_t *ppos)
 229 {
 230         struct hstate *h = hstate_file(filp);
 231         struct address_space *mapping = filp->f_mapping;
 232         struct inode *inode = mapping->host;
 233         unsigned long index = *ppos >> huge_page_shift(h);
 234         unsigned long offset = *ppos & ~huge_page_mask(h);
 235         unsigned long end_index;
 236         loff_t isize;
 237         ssize_t retval = 0;
 238
 239         /* validate length */
 240         if (len == 0)
 241                 goto out;
 242
 243         for (;;) {
 244                 struct page *page;
 245                 unsigned long nr, ret;
 246                 int ra;
 247
 248                 /* nr is the maximum number of bytes to copy from this page */
 249                 nr = huge_page_size(h);
 250                 isize = i_size_read(inode);
 251                 if (!isize)
 252                         goto out;
 253                 end_index = (isize - 1) >> huge_page_shift(h);
 254                 if (index >= end_index) {
 255                         if (index > end_index)
 256                                 goto out;
 257                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 258                         if (nr <= offset)
 259                                 goto out;
 260                 }
 261                 nr = nr - offset;
 262
 263                 /* Find the page */
 264                 page = find_lock_page(mapping, index);
 265                 if (unlikely(page == NULL)) {
 266                         /*
 267                          * We have a HOLE, zero out the user-buffer for the
 268                          * length of the hole or request.
 269                          */
 270                         ret = len < nr ? len : nr;
 271                         if (clear_user(buf, ret))
 272                                 ra = -EFAULT;
 273                         else
 274                                 ra = 0;
 275                 } else {
 276                         unlock_page(page);
 277
 278                         /*
 279                          * We have the page, copy it to user space buffer.
 280                          */
 281                         ra = hugetlbfs_read_actor(page, offset, buf, len, nr);
 282                         ret = ra;
 283                         page_cache_release(page);
 284                 }
 285                 if (ra < 0) {
 286                         if (retval == 0)
 287                                 retval = ra;
 288                         goto out;
 289                 }
 290
 291                 offset += ret;
 292                 retval += ret;
 293                 len -= ret;
 294                 index += offset >> huge_page_shift(h);
 295                 offset &= ~huge_page_mask(h);
 296
 297                 /* short read or no more work */
 298                 if ((ret != nr) || (len == 0))
 299                         break;
 300         }
 301 out:
 302         *ppos = ((loff_t)index << huge_page_shift(h)) + offset;
 303         return retval;
 304 }
 305
 306 static int hugetlbfs_write_begin(struct file *file,
 307                         struct address_space *mapping,
 308                         loff_t pos, unsigned len, unsigned flags,
 309                         struct page **pagep, void **fsdata)
 310 {
 311         return -EINVAL;
 312 }
 313
 314 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 315                         loff_t pos, unsigned len, unsigned copied,
 316                         struct page *page, void *fsdata)
 317 {
 318         BUG();
 319         return -EINVAL;
 320 }
 321
 322 static void truncate_huge_page(struct page *page)
 323 {
 324         ClearPageDirty(page);
 325         ClearPageUptodate(page);
 326         delete_from_page_cache(page);
 327 }
 328
 329 static void truncate_hugepages(struct inode *inode, loff_t lstart)
 330 {
 331         struct hstate *h = hstate_inode(inode);
 332         struct address_space *mapping = &inode->i_data;
 333         const pgoff_t start = lstart >> huge_page_shift(h);
 334         struct pagevec pvec;
 335         pgoff_t next;
 336         int i, freed = 0;
 337
 338         pagevec_init(&pvec, 0);
 339         next = start;
 340         while (1) {
 341                 if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
 342                         if (next == start)
 343                                 break;
 344                         next = start;
 345                         continue;
 346                 }
 347
 348                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 349                         struct page *page = pvec.pages[i];
 350
 351                         lock_page(page);
 352                         if (page->index > next)
 353                                 next = page->index;
 354                         ++next;
 355                         truncate_huge_page(page);
 356                         unlock_page(page);
 357                         freed++;
 358                 }
 359                 huge_pagevec_release(&pvec);
 360         }
 361         BUG_ON(!lstart && mapping->nrpages);
 362         hugetlb_unreserve_pages(inode, start, freed);
 363 }
 364
 365 static void hugetlbfs_evict_inode(struct inode *inode)
 366 {
 367         struct resv_map *resv_map;
 368
 369         truncate_hugepages(inode, 0);
 370         resv_map = (struct resv_map *)inode->i_mapping->private_data;
 371         /* root inode doesn't have the resv_map, so we should check it */
 372         if (resv_map)
 373                 resv_map_release(&resv_map->refs);
 374         clear_inode(inode);
 375 }
 376
 377 static inline void
 378 hugetlb_vmtruncate_list(struct rb_root *root, pgoff_t pgoff)
 379 {
 380         struct vm_area_struct *vma;
 381
 382         vma_interval_tree_foreach(vma, root, pgoff, ULONG_MAX) {
 383                 unsigned long v_offset;
 384
 385                 /*
 386                  * Can the expression below overflow on 32-bit arches?
 387                  * No, because the interval tree returns us only those vmas
 388                  * which overlap the truncated area starting at pgoff,
 389                  * and no vma on a 32-bit arch can span beyond the 4GB.
 390                  */
 391                 if (vma->vm_pgoff < pgoff)
 392                         v_offset = (pgoff - vma->vm_pgoff) << PAGE_SHIFT;
 393                 else
 394                         v_offset = 0;
 395
 396                 unmap_hugepage_range(vma, vma->vm_start + v_offset,
 397                                      vma->vm_end, NULL);
 398         }
 399 }
 400
 401 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 402 {
 403         pgoff_t pgoff;
 404         struct address_space *mapping = inode->i_mapping;
 405         struct hstate *h = hstate_inode(inode);
 406
 407         BUG_ON(offset & ~huge_page_mask(h));
 408         pgoff = offset >> PAGE_SHIFT;
 409
 410         i_size_write(inode, offset);
 411         i_mmap_lock_write(mapping);
 412         if (!RB_EMPTY_ROOT(&mapping->i_mmap))
 413                 hugetlb_vmtruncate_list(&mapping->i_mmap, pgoff);
 414         i_mmap_unlock_write(mapping);
 415         truncate_hugepages(inode, offset);
 416         return 0;
 417 }
 418
 419 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 420 {
 421         struct inode *inode = dentry->d_inode;
 422         struct hstate *h = hstate_inode(inode);
 423         int error;
 424         unsigned int ia_valid = attr->ia_valid;
 425
 426         BUG_ON(!inode);
 427
 428         error = inode_change_ok(inode, attr);
 429         if (error)
 430                 return error;
 431
 432         if (ia_valid & ATTR_SIZE) {
 433                 error = -EINVAL;
 434                 if (attr->ia_size & ~huge_page_mask(h))
 435                         return -EINVAL;
 436                 error = hugetlb_vmtruncate(inode, attr->ia_size);
 437                 if (error)
 438                         return error;
 439         }
 440
 441         setattr_copy(inode, attr);
 442         mark_inode_dirty(inode);
 443         return 0;
 444 }
 445
 446 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 447                                         struct hugetlbfs_config *config)
 448 {
 449         struct inode *inode;
 450
 451         inode = new_inode(sb);
 452         if (inode) {
 453                 struct hugetlbfs_inode_info *info;
 454                 inode->i_ino = get_next_ino();
 455                 inode->i_mode = S_IFDIR | config->mode;
 456                 inode->i_uid = config->uid;
 457                 inode->i_gid = config->gid;
 458                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 459                 info = HUGETLBFS_I(inode);
 460                 mpol_shared_policy_init(&info->policy, NULL);
 461                 inode->i_op = &hugetlbfs_dir_inode_operations;
 462                 inode->i_fop = &simple_dir_operations;
 463                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 464                 inc_nlink(inode);
 465                 lockdep_annotate_inode_mutex_key(inode);
 466         }
 467         return inode;
 468 }
 469
 470 /*
 471  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 472  * be taken from reclaim -- unlike regular filesystems. This needs an
 473  * annotation because huge_pmd_share() does an allocation under
 474  * i_mmap_rwsem.
 475  */
 476 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 477
 478 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 479                                         struct inode *dir,
 480                                         umode_t mode, dev_t dev)
 481 {
 482         struct inode *inode;
 483         struct resv_map *resv_map;
 484
 485         resv_map = resv_map_alloc();
 486         if (!resv_map)
 487                 return NULL;
 488
 489         inode = new_inode(sb);
 490         if (inode) {
 491                 struct hugetlbfs_inode_info *info;
 492                 inode->i_ino = get_next_ino();
 493                 inode_init_owner(inode, dir, mode);
 494                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 495                                 &hugetlbfs_i_mmap_rwsem_key);
 496                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 497                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 498                 inode->i_mapping->private_data = resv_map;
 499                 info = HUGETLBFS_I(inode);
 500                 /*
 501                  * The policy is initialized here even if we are creating a
 502                  * private inode because initialization simply creates an
 503                  * an empty rb tree and calls spin_lock_init(), later when we
 504                  * call mpol_free_shared_policy() it will just return because
 505                  * the rb tree will still be empty.
 506                  */
 507                 mpol_shared_policy_init(&info->policy, NULL);
 508                 switch (mode & S_IFMT) {
 509                 default:
 510                         init_special_inode(inode, mode, dev);
 511                         break;
 512                 case S_IFREG:
 513                         inode->i_op = &hugetlbfs_inode_operations;
 514                         inode->i_fop = &hugetlbfs_file_operations;
 515                         break;
 516                 case S_IFDIR:
 517                         inode->i_op = &hugetlbfs_dir_inode_operations;
 518                         inode->i_fop = &simple_dir_operations;
 519
 520                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 521                         inc_nlink(inode);
 522                         break;
 523                 case S_IFLNK:
 524                         inode->i_op = &page_symlink_inode_operations;
 525                         break;
 526                 }
 527                 lockdep_annotate_inode_mutex_key(inode);
 528         } else
 529                 kref_put(&resv_map->refs, resv_map_release);
 530
 531         return inode;
 532 }
 533
 534 /*
 535  * File creation. Allocate an inode, and we're done..
 536  */
 537 static int hugetlbfs_mknod(struct inode *dir,
 538                         struct dentry *dentry, umode_t mode, dev_t dev)
 539 {
 540         struct inode *inode;
 541         int error = -ENOSPC;
 542
 543         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 544         if (inode) {
 545                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 546                 d_instantiate(dentry, inode);
 547                 dget(dentry);   /* Extra count - pin the dentry in core */
 548                 error = 0;
 549         }
 550         return error;
 551 }
 552
 553 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 554 {
 555         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 556         if (!retval)
 557                 inc_nlink(dir);
 558         return retval;
 559 }
 560
 561 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 562 {
 563         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 564 }
 565
 566 static int hugetlbfs_symlink(struct inode *dir,
 567                         struct dentry *dentry, const char *symname)
 568 {
 569         struct inode *inode;
 570         int error = -ENOSPC;
 571
 572         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 573         if (inode) {
 574                 int l = strlen(symname)+1;
 575                 error = page_symlink(inode, symname, l);
 576                 if (!error) {
 577                         d_instantiate(dentry, inode);
 578                         dget(dentry);
 579                 } else
 580                         iput(inode);
 581         }
 582         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 583
 584         return error;
 585 }
 586
 587 /*
 588  * mark the head page dirty
 589  */
 590 static int hugetlbfs_set_page_dirty(struct page *page)
 591 {
 592         struct page *head = compound_head(page);
 593
 594         SetPageDirty(head);
 595         return 0;
 596 }
 597
 598 static int hugetlbfs_migrate_page(struct address_space *mapping,
 599                                 struct page *newpage, struct page *page,
 600                                 enum migrate_mode mode)
 601 {
 602         int rc;
 603
 604         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 605         if (rc != MIGRATEPAGE_SUCCESS)
 606                 return rc;
 607         migrate_page_copy(newpage, page);
 608
 609         return MIGRATEPAGE_SUCCESS;
 610 }
 611
 612 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 613 {
 614         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 615         struct hstate *h = hstate_inode(dentry->d_inode);
 616
 617         buf->f_type = HUGETLBFS_MAGIC;
 618         buf->f_bsize = huge_page_size(h);
 619         if (sbinfo) {
 620                 spin_lock(&sbinfo->stat_lock);
 621                 /* If no limits set, just report 0 for max/free/used
 622                  * blocks, like simple_statfs() */
 623                 if (sbinfo->spool) {
 624                         long free_pages;
 625
 626                         spin_lock(&sbinfo->spool->lock);
 627                         buf->f_blocks = sbinfo->spool->max_hpages;
 628                         free_pages = sbinfo->spool->max_hpages
 629                                 - sbinfo->spool->used_hpages;
 630                         buf->f_bavail = buf->f_bfree = free_pages;
 631                         spin_unlock(&sbinfo->spool->lock);
 632                         buf->f_files = sbinfo->max_inodes;
 633                         buf->f_ffree = sbinfo->free_inodes;
 634                 }
 635                 spin_unlock(&sbinfo->stat_lock);
 636         }
 637         buf->f_namelen = NAME_MAX;
 638         return 0;
 639 }
 640
 641 static void hugetlbfs_put_super(struct super_block *sb)
 642 {
 643         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
 644
 645         if (sbi) {
 646                 sb->s_fs_info = NULL;
 647
 648                 if (sbi->spool)
 649                         hugepage_put_subpool(sbi->spool);
 650
 651                 kfree(sbi);
 652         }
 653 }
 654
 655 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 656 {
 657         if (sbinfo->free_inodes >= 0) {
 658                 spin_lock(&sbinfo->stat_lock);
 659                 if (unlikely(!sbinfo->free_inodes)) {
 660                         spin_unlock(&sbinfo->stat_lock);
 661                         return 0;
 662                 }
 663                 sbinfo->free_inodes--;
 664                 spin_unlock(&sbinfo->stat_lock);
 665         }
 666
 667         return 1;
 668 }
 669
 670 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 671 {
 672         if (sbinfo->free_inodes >= 0) {
 673                 spin_lock(&sbinfo->stat_lock);
 674                 sbinfo->free_inodes++;
 675                 spin_unlock(&sbinfo->stat_lock);
 676         }
 677 }
 678
 679
 680 static struct kmem_cache *hugetlbfs_inode_cachep;
 681
 682 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
 683 {
 684         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
 685         struct hugetlbfs_inode_info *p;
 686
 687         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
 688                 return NULL;
 689         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
 690         if (unlikely(!p)) {
 691                 hugetlbfs_inc_free_inodes(sbinfo);
 692                 return NULL;
 693         }
 694         return &p->vfs_inode;
 695 }
 696
 697 static void hugetlbfs_i_callback(struct rcu_head *head)
 698 {
 699         struct inode *inode = container_of(head, struct inode, i_rcu);
 700         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
 701 }
 702
 703 static void hugetlbfs_destroy_inode(struct inode *inode)
 704 {
 705         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
 706         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
 707         call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
 708 }
 709
 710 static const struct address_space_operations hugetlbfs_aops = {
 711         .write_begin    = hugetlbfs_write_begin,
 712         .write_end      = hugetlbfs_write_end,
 713         .set_page_dirty = hugetlbfs_set_page_dirty,
 714         .migratepage    = hugetlbfs_migrate_page,
 715 };
 716
 717
 718 static void init_once(void *foo)
 719 {
 720         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
 721
 722         inode_init_once(&ei->vfs_inode);
 723 }
 724
 725 const struct file_operations hugetlbfs_file_operations = {
 726         .read                   = hugetlbfs_read,
 727         .mmap                   = hugetlbfs_file_mmap,
 728         .fsync                  = noop_fsync,
 729         .get_unmapped_area      = hugetlb_get_unmapped_area,
 730         .llseek         = default_llseek,
 731 };
 732
 733 static const struct inode_operations hugetlbfs_dir_inode_operations = {
 734         .create         = hugetlbfs_create,
 735         .lookup         = simple_lookup,
 736         .link           = simple_link,
 737         .unlink         = simple_unlink,
 738         .symlink        = hugetlbfs_symlink,
 739         .mkdir          = hugetlbfs_mkdir,
 740         .rmdir          = simple_rmdir,
 741         .mknod          = hugetlbfs_mknod,
 742         .rename         = simple_rename,
 743         .setattr        = hugetlbfs_setattr,
 744 };
 745
 746 static const struct inode_operations hugetlbfs_inode_operations = {
 747         .setattr        = hugetlbfs_setattr,
 748 };
 749
 750 static const struct super_operations hugetlbfs_ops = {
 751         .alloc_inode    = hugetlbfs_alloc_inode,
 752         .destroy_inode  = hugetlbfs_destroy_inode,
 753         .evict_inode    = hugetlbfs_evict_inode,
 754         .statfs         = hugetlbfs_statfs,
 755         .put_super      = hugetlbfs_put_super,
 756         .show_options   = generic_show_options,
 757 };
 758
 759 enum { NO_SIZE, SIZE_STD, SIZE_PERCENT };
 760
 761 /*
 762  * Convert size option passed from command line to number of huge pages
 763  * in the pool specified by hstate.  Size option could be in bytes
 764  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
 765  */
 766 static long long
 767 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
 768                                                                 int val_type)
 769 {
 770         if (val_type == NO_SIZE)
 771                 return -1;
 772
 773         if (val_type == SIZE_PERCENT) {
 774                 size_opt <<= huge_page_shift(h);
 775                 size_opt *= h->max_huge_pages;
 776                 do_div(size_opt, 100);
 777         }
 778
 779         size_opt >>= huge_page_shift(h);
 780         return size_opt;
 781 }
 782
 783 static int
 784 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
 785 {
 786         char *p, *rest;
 787         substring_t args[MAX_OPT_ARGS];
 788         int option;
 789         unsigned long long max_size_opt = 0, min_size_opt = 0;
 790         int max_val_type = NO_SIZE, min_val_type = NO_SIZE;
 791
 792         if (!options)
 793                 return 0;
 794
 795         while ((p = strsep(&options, ",")) != NULL) {
 796                 int token;
 797                 if (!*p)
 798                         continue;
 799
 800                 token = match_token(p, tokens, args);
 801                 switch (token) {
 802                 case Opt_uid:
 803                         if (match_int(&args[0], &option))
 804                                 goto bad_val;
 805                         pconfig->uid = make_kuid(current_user_ns(), option);
 806                         if (!uid_valid(pconfig->uid))
 807                                 goto bad_val;
 808                         break;
 809
 810                 case Opt_gid:
 811                         if (match_int(&args[0], &option))
 812                                 goto bad_val;
 813                         pconfig->gid = make_kgid(current_user_ns(), option);
 814                         if (!gid_valid(pconfig->gid))
 815                                 goto bad_val;
 816                         break;
 817
 818                 case Opt_mode:
 819                         if (match_octal(&args[0], &option))
 820                                 goto bad_val;
 821                         pconfig->mode = option & 01777U;
 822                         break;
 823
 824                 case Opt_size: {
 825                         /* memparse() will accept a K/M/G without a digit */
 826                         if (!isdigit(*args[0].from))
 827                                 goto bad_val;
 828                         max_size_opt = memparse(args[0].from, &rest);
 829                         max_val_type = SIZE_STD;
 830                         if (*rest == '%')
 831                                 max_val_type = SIZE_PERCENT;
 832                         break;
 833                 }
 834
 835                 case Opt_nr_inodes:
 836                         /* memparse() will accept a K/M/G without a digit */
 837                         if (!isdigit(*args[0].from))
 838                                 goto bad_val;
 839                         pconfig->nr_inodes = memparse(args[0].from, &rest);
 840                         break;
 841
 842                 case Opt_pagesize: {
 843                         unsigned long ps;
 844                         ps = memparse(args[0].from, &rest);
 845                         pconfig->hstate = size_to_hstate(ps);
 846                         if (!pconfig->hstate) {
 847                                 pr_err("Unsupported page size %lu MB\n",
 848                                         ps >> 20);
 849                                 return -EINVAL;
 850                         }
 851                         break;
 852                 }
 853
 854                 case Opt_min_size: {
 855                         /* memparse() will accept a K/M/G without a digit */
 856                         if (!isdigit(*args[0].from))
 857                                 goto bad_val;
 858                         min_size_opt = memparse(args[0].from, &rest);
 859                         min_val_type = SIZE_STD;
 860                         if (*rest == '%')
 861                                 min_val_type = SIZE_PERCENT;
 862                         break;
 863                 }
 864
 865                 default:
 866                         pr_err("Bad mount option: \"%s\"\n", p);
 867                         return -EINVAL;
 868                         break;
 869                 }
 870         }
 871
 872         /*
 873          * Use huge page pool size (in hstate) to convert the size
 874          * options to number of huge pages.  If NO_SIZE, -1 is returned.
 875          */
 876         pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
 877                                                 max_size_opt, max_val_type);
 878         pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
 879                                                 min_size_opt, min_val_type);
 880
 881         /*
 882          * If max_size was specified, then min_size must be smaller
 883          */
 884         if (max_val_type > NO_SIZE &&
 885             pconfig->min_hpages > pconfig->max_hpages) {
 886                 pr_err("minimum size can not be greater than maximum size\n");
 887                 return -EINVAL;
 888         }
 889
 890         return 0;
 891
 892 bad_val:
 893         pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
 894         return -EINVAL;
 895 }
 896
 897 static int
 898 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
 899 {
 900         int ret;
 901         struct hugetlbfs_config config;
 902         struct hugetlbfs_sb_info *sbinfo;
 903
 904         save_mount_options(sb, data);
 905
 906         config.max_hpages = -1; /* No limit on size by default */
 907         config.nr_inodes = -1; /* No limit on number of inodes by default */
 908         config.uid = current_fsuid();
 909         config.gid = current_fsgid();
 910         config.mode = 0755;
 911         config.hstate = &default_hstate;
 912         config.min_hpages = -1; /* No default minimum size */
 913         ret = hugetlbfs_parse_options(data, &config);
 914         if (ret)
 915                 return ret;
 916
 917         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
 918         if (!sbinfo)
 919                 return -ENOMEM;
 920         sb->s_fs_info = sbinfo;
 921         sbinfo->hstate = config.hstate;
 922         spin_lock_init(&sbinfo->stat_lock);
 923         sbinfo->max_inodes = config.nr_inodes;
 924         sbinfo->free_inodes = config.nr_inodes;
 925         sbinfo->spool = NULL;
 926         /*
 927          * Allocate and initialize subpool if maximum or minimum size is
 928          * specified.  Any needed reservations (for minimim size) are taken
 929          * taken when the subpool is created.
 930          */
 931         if (config.max_hpages != -1 || config.min_hpages != -1) {
 932                 sbinfo->spool = hugepage_new_subpool(config.hstate,
 933                                                         config.max_hpages,
 934                                                         config.min_hpages);
 935                 if (!sbinfo->spool)
 936                         goto out_free;
 937         }
 938         sb->s_maxbytes = MAX_LFS_FILESIZE;
 939         sb->s_blocksize = huge_page_size(config.hstate);
 940         sb->s_blocksize_bits = huge_page_shift(config.hstate);
 941         sb->s_magic = HUGETLBFS_MAGIC;
 942         sb->s_op = &hugetlbfs_ops;
 943         sb->s_time_gran = 1;
 944         sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
 945         if (!sb->s_root)
 946                 goto out_free;
 947         return 0;
 948 out_free:
 949         kfree(sbinfo->spool);
 950         kfree(sbinfo);
 951         return -ENOMEM;
 952 }
 953
 954 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
 955         int flags, const char *dev_name, void *data)
 956 {
 957         return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
 958 }
 959
 960 static struct file_system_type hugetlbfs_fs_type = {
 961         .name           = "hugetlbfs",
 962         .mount          = hugetlbfs_mount,
 963         .kill_sb        = kill_litter_super,
 964 };
 965 MODULE_ALIAS_FS("hugetlbfs");
 966
 967 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
 968
 969 static int can_do_hugetlb_shm(void)
 970 {
 971         kgid_t shm_group;
 972         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
 973         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
 974 }
 975
 976 static int get_hstate_idx(int page_size_log)
 977 {
 978         struct hstate *h = hstate_sizelog(page_size_log);
 979
 980         if (!h)
 981                 return -1;
 982         return h - hstates;
 983 }
 984
 985 static const struct dentry_operations anon_ops = {
 986         .d_dname = simple_dname
 987 };
 988
 989 /*
 990  * Note that size should be aligned to proper hugepage size in caller side,
 991  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
 992  */
 993 struct file *hugetlb_file_setup(const char *name, size_t size,
 994                                 vm_flags_t acctflag, struct user_struct **user,
 995                                 int creat_flags, int page_size_log)
 996 {
 997         struct file *file = ERR_PTR(-ENOMEM);
 998         struct inode *inode;
 999         struct path path;
1000         struct super_block *sb;
1001         struct qstr quick_string;
1002         int hstate_idx;
1003
1004         hstate_idx = get_hstate_idx(page_size_log);
1005         if (hstate_idx < 0)
1006                 return ERR_PTR(-ENODEV);
1007
1008         *user = NULL;
1009         if (!hugetlbfs_vfsmount[hstate_idx])
1010                 return ERR_PTR(-ENOENT);
1011
1012         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1013                 *user = current_user();
1014                 if (user_shm_lock(size, *user)) {
1015                         task_lock(current);
1016                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1017                                 current->comm, current->pid);
1018                         task_unlock(current);
1019                 } else {
1020                         *user = NULL;
1021                         return ERR_PTR(-EPERM);
1022                 }
1023         }
1024
1025         sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
1026         quick_string.name = name;
1027         quick_string.len = strlen(quick_string.name);
1028         quick_string.hash = 0;
1029         path.dentry = d_alloc_pseudo(sb, &quick_string);
1030         if (!path.dentry)
1031                 goto out_shm_unlock;
1032
1033         d_set_d_op(path.dentry, &anon_ops);
1034         path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
1035         file = ERR_PTR(-ENOSPC);
1036         inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
1037         if (!inode)
1038                 goto out_dentry;
1039
1040         file = ERR_PTR(-ENOMEM);
1041         if (hugetlb_reserve_pages(inode, 0,
1042                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1043                         acctflag))
1044                 goto out_inode;
1045
1046         d_instantiate(path.dentry, inode);
1047         inode->i_size = size;
1048         clear_nlink(inode);
1049
1050         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1051                         &hugetlbfs_file_operations);
1052         if (IS_ERR(file))
1053                 goto out_dentry; /* inode is already attached */
1054
1055         return file;
1056
1057 out_inode:
1058         iput(inode);
1059 out_dentry:
1060         path_put(&path);
1061 out_shm_unlock:
1062         if (*user) {
1063                 user_shm_unlock(size, *user);
1064                 *user = NULL;
1065         }
1066         return file;
1067 }
1068
1069 static int __init init_hugetlbfs_fs(void)
1070 {
1071         struct hstate *h;
1072         int error;
1073         int i;
1074
1075         if (!hugepages_supported()) {
1076                 pr_info("disabling because there are no supported hugepage sizes\n");
1077                 return -ENOTSUPP;
1078         }
1079
1080         error = -ENOMEM;
1081         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1082                                         sizeof(struct hugetlbfs_inode_info),
1083                                         0, 0, init_once);
1084         if (hugetlbfs_inode_cachep == NULL)
1085                 goto out2;
1086
1087         error = register_filesystem(&hugetlbfs_fs_type);
1088         if (error)
1089                 goto out;
1090
1091         i = 0;
1092         for_each_hstate(h) {
1093                 char buf[50];
1094                 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1095
1096                 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1097                 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1098                                                         buf);
1099
1100                 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1101                         pr_err("Cannot mount internal hugetlbfs for "
1102                                 "page size %uK", ps_kb);
1103                         error = PTR_ERR(hugetlbfs_vfsmount[i]);
1104                         hugetlbfs_vfsmount[i] = NULL;
1105                 }
1106                 i++;
1107         }
1108         /* Non default hstates are optional */
1109         if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1110                 return 0;
1111
1112  out:
1113         kmem_cache_destroy(hugetlbfs_inode_cachep);
1114  out2:
1115         return error;
1116 }
1117
1118 static void __exit exit_hugetlbfs_fs(void)
1119 {
1120         struct hstate *h;
1121         int i;
1122
1123
1124         /*
1125          * Make sure all delayed rcu free inodes are flushed before we
1126          * destroy cache.
1127          */
1128         rcu_barrier();
1129         kmem_cache_destroy(hugetlbfs_inode_cachep);
1130         i = 0;
1131         for_each_hstate(h)
1132                 kern_unmount(hugetlbfs_vfsmount[i++]);
1133         unregister_filesystem(&hugetlbfs_fs_type);
1134 }
1135
1136 module_init(init_hugetlbfs_fs)
1137 module_exit(exit_hugetlbfs_fs)
1138
1139 MODULE_LICENSE("GPL");