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  * License: GPL
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/thread_info.h>
  13 #include <asm/current.h>
  14 #include <linux/falloc.h>
  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/fs_parser.h>
  30 #include <linux/mman.h>
  31 #include <linux/slab.h>
  32 #include <linux/dnotify.h>
  33 #include <linux/statfs.h>
  34 #include <linux/security.h>
  35 #include <linux/magic.h>
  36 #include <linux/migrate.h>
  37 #include <linux/uio.h>
  38
  39 #include <linux/uaccess.h>
  40 #include <linux/sched/mm.h>
  41
  42 static const struct address_space_operations hugetlbfs_aops;
  43 const struct file_operations hugetlbfs_file_operations;
  44 static const struct inode_operations hugetlbfs_dir_inode_operations;
  45 static const struct inode_operations hugetlbfs_inode_operations;
  46
  47 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
  48
  49 struct hugetlbfs_fs_context {
  50         struct hstate           *hstate;
  51         unsigned long long      max_size_opt;
  52         unsigned long long      min_size_opt;
  53         long                    max_hpages;
  54         long                    nr_inodes;
  55         long                    min_hpages;
  56         enum hugetlbfs_size_type max_val_type;
  57         enum hugetlbfs_size_type min_val_type;
  58         kuid_t                  uid;
  59         kgid_t                  gid;
  60         umode_t                 mode;
  61 };
  62
  63 int sysctl_hugetlb_shm_group;
  64
  65 enum hugetlb_param {
  66         Opt_gid,
  67         Opt_min_size,
  68         Opt_mode,
  69         Opt_nr_inodes,
  70         Opt_pagesize,
  71         Opt_size,
  72         Opt_uid,
  73 };
  74
  75 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
  76         fsparam_u32   ("gid",           Opt_gid),
  77         fsparam_string("min_size",      Opt_min_size),
  78         fsparam_u32oct("mode",          Opt_mode),
  79         fsparam_string("nr_inodes",     Opt_nr_inodes),
  80         fsparam_string("pagesize",      Opt_pagesize),
  81         fsparam_string("size",          Opt_size),
  82         fsparam_u32   ("uid",           Opt_uid),
  83         {}
  84 };
  85
  86 #ifdef CONFIG_NUMA
  87 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  88                                         struct inode *inode, pgoff_t index)
  89 {
  90         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  91                                                         index);
  92 }
  93
  94 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  95 {
  96         mpol_cond_put(vma->vm_policy);
  97 }
  98 #else
  99 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
 100                                         struct inode *inode, pgoff_t index)
 101 {
 102 }
 103
 104 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
 105 {
 106 }
 107 #endif
 108
 109 /*
 110  * Mask used when checking the page offset value passed in via system
 111  * calls.  This value will be converted to a loff_t which is signed.
 112  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 113  * value.  The extra bit (- 1 in the shift value) is to take the sign
 114  * bit into account.
 115  */
 116 #define PGOFF_LOFFT_MAX \
 117         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 118
 119 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 120 {
 121         struct inode *inode = file_inode(file);
 122         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 123         loff_t len, vma_len;
 124         int ret;
 125         struct hstate *h = hstate_file(file);
 126
 127         /*
 128          * vma address alignment (but not the pgoff alignment) has
 129          * already been checked by prepare_hugepage_range.  If you add
 130          * any error returns here, do so after setting VM_HUGETLB, so
 131          * is_vm_hugetlb_page tests below unmap_region go the right
 132          * way when do_mmap unwinds (may be important on powerpc
 133          * and ia64).
 134          */
 135         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 136         vma->vm_ops = &hugetlb_vm_ops;
 137
 138         ret = seal_check_future_write(info->seals, vma);
 139         if (ret)
 140                 return ret;
 141
 142         /*
 143          * page based offset in vm_pgoff could be sufficiently large to
 144          * overflow a loff_t when converted to byte offset.  This can
 145          * only happen on architectures where sizeof(loff_t) ==
 146          * sizeof(unsigned long).  So, only check in those instances.
 147          */
 148         if (sizeof(unsigned long) == sizeof(loff_t)) {
 149                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 150                         return -EINVAL;
 151         }
 152
 153         /* must be huge page aligned */
 154         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 155                 return -EINVAL;
 156
 157         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 158         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 159         /* check for overflow */
 160         if (len < vma_len)
 161                 return -EINVAL;
 162
 163         inode_lock(inode);
 164         file_accessed(file);
 165
 166         ret = -ENOMEM;
 167         if (!hugetlb_reserve_pages(inode,
 168                                 vma->vm_pgoff >> huge_page_order(h),
 169                                 len >> huge_page_shift(h), vma,
 170                                 vma->vm_flags))
 171                 goto out;
 172
 173         ret = 0;
 174         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 175                 i_size_write(inode, len);
 176 out:
 177         inode_unlock(inode);
 178
 179         return ret;
 180 }
 181
 182 /*
 183  * Called under mmap_write_lock(mm).
 184  */
 185
 186 static unsigned long
 187 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
 188                 unsigned long len, unsigned long pgoff, unsigned long flags)
 189 {
 190         struct hstate *h = hstate_file(file);
 191         struct vm_unmapped_area_info info;
 192
 193         info.flags = 0;
 194         info.length = len;
 195         info.low_limit = current->mm->mmap_base;
 196         info.high_limit = arch_get_mmap_end(addr, len, flags);
 197         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 198         info.align_offset = 0;
 199         return vm_unmapped_area(&info);
 200 }
 201
 202 static unsigned long
 203 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
 204                 unsigned long len, unsigned long pgoff, unsigned long flags)
 205 {
 206         struct hstate *h = hstate_file(file);
 207         struct vm_unmapped_area_info info;
 208
 209         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 210         info.length = len;
 211         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
 212         info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
 213         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 214         info.align_offset = 0;
 215         addr = vm_unmapped_area(&info);
 216
 217         /*
 218          * A failed mmap() very likely causes application failure,
 219          * so fall back to the bottom-up function here. This scenario
 220          * can happen with large stack limits and large mmap()
 221          * allocations.
 222          */
 223         if (unlikely(offset_in_page(addr))) {
 224                 VM_BUG_ON(addr != -ENOMEM);
 225                 info.flags = 0;
 226                 info.low_limit = current->mm->mmap_base;
 227                 info.high_limit = arch_get_mmap_end(addr, len, flags);
 228                 addr = vm_unmapped_area(&info);
 229         }
 230
 231         return addr;
 232 }
 233
 234 unsigned long
 235 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 236                                   unsigned long len, unsigned long pgoff,
 237                                   unsigned long flags)
 238 {
 239         struct mm_struct *mm = current->mm;
 240         struct vm_area_struct *vma;
 241         struct hstate *h = hstate_file(file);
 242         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
 243
 244         if (len & ~huge_page_mask(h))
 245                 return -EINVAL;
 246         if (len > TASK_SIZE)
 247                 return -ENOMEM;
 248
 249         if (flags & MAP_FIXED) {
 250                 if (prepare_hugepage_range(file, addr, len))
 251                         return -EINVAL;
 252                 return addr;
 253         }
 254
 255         if (addr) {
 256                 addr = ALIGN(addr, huge_page_size(h));
 257                 vma = find_vma(mm, addr);
 258                 if (mmap_end - len >= addr &&
 259                     (!vma || addr + len <= vm_start_gap(vma)))
 260                         return addr;
 261         }
 262
 263         /*
 264          * Use mm->get_unmapped_area value as a hint to use topdown routine.
 265          * If architectures have special needs, they should define their own
 266          * version of hugetlb_get_unmapped_area.
 267          */
 268         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
 269                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 270                                 pgoff, flags);
 271         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 272                         pgoff, flags);
 273 }
 274
 275 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 276 static unsigned long
 277 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 278                           unsigned long len, unsigned long pgoff,
 279                           unsigned long flags)
 280 {
 281         return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
 282 }
 283 #endif
 284
 285 /*
 286  * Support for read() - Find the page attached to f_mapping and copy out the
 287  * data. This provides functionality similar to filemap_read().
 288  */
 289 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 290 {
 291         struct file *file = iocb->ki_filp;
 292         struct hstate *h = hstate_file(file);
 293         struct address_space *mapping = file->f_mapping;
 294         struct inode *inode = mapping->host;
 295         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 296         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 297         unsigned long end_index;
 298         loff_t isize;
 299         ssize_t retval = 0;
 300
 301         while (iov_iter_count(to)) {
 302                 struct page *page;
 303                 size_t nr, copied;
 304
 305                 /* nr is the maximum number of bytes to copy from this page */
 306                 nr = huge_page_size(h);
 307                 isize = i_size_read(inode);
 308                 if (!isize)
 309                         break;
 310                 end_index = (isize - 1) >> huge_page_shift(h);
 311                 if (index > end_index)
 312                         break;
 313                 if (index == end_index) {
 314                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 315                         if (nr <= offset)
 316                                 break;
 317                 }
 318                 nr = nr - offset;
 319
 320                 /* Find the page */
 321                 page = find_lock_page(mapping, index);
 322                 if (unlikely(page == NULL)) {
 323                         /*
 324                          * We have a HOLE, zero out the user-buffer for the
 325                          * length of the hole or request.
 326                          */
 327                         copied = iov_iter_zero(nr, to);
 328                 } else {
 329                         unlock_page(page);
 330
 331                         if (PageHWPoison(page)) {
 332                                 put_page(page);
 333                                 retval = -EIO;
 334                                 break;
 335                         }
 336
 337                         /*
 338                          * We have the page, copy it to user space buffer.
 339                          */
 340                         copied = copy_page_to_iter(page, offset, nr, to);
 341                         put_page(page);
 342                 }
 343                 offset += copied;
 344                 retval += copied;
 345                 if (copied != nr && iov_iter_count(to)) {
 346                         if (!retval)
 347                                 retval = -EFAULT;
 348                         break;
 349                 }
 350                 index += offset >> huge_page_shift(h);
 351                 offset &= ~huge_page_mask(h);
 352         }
 353         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 354         return retval;
 355 }
 356
 357 static int hugetlbfs_write_begin(struct file *file,
 358                         struct address_space *mapping,
 359                         loff_t pos, unsigned len,
 360                         struct page **pagep, void **fsdata)
 361 {
 362         return -EINVAL;
 363 }
 364
 365 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 366                         loff_t pos, unsigned len, unsigned copied,
 367                         struct page *page, void *fsdata)
 368 {
 369         BUG();
 370         return -EINVAL;
 371 }
 372
 373 static void hugetlb_delete_from_page_cache(struct folio *folio)
 374 {
 375         folio_clear_dirty(folio);
 376         folio_clear_uptodate(folio);
 377         filemap_remove_folio(folio);
 378 }
 379
 380 /*
 381  * Called with i_mmap_rwsem held for inode based vma maps.  This makes
 382  * sure vma (and vm_mm) will not go away.  We also hold the hugetlb fault
 383  * mutex for the page in the mapping.  So, we can not race with page being
 384  * faulted into the vma.
 385  */
 386 static bool hugetlb_vma_maps_page(struct vm_area_struct *vma,
 387                                 unsigned long addr, struct page *page)
 388 {
 389         pte_t *ptep, pte;
 390
 391         ptep = huge_pte_offset(vma->vm_mm, addr,
 392                         huge_page_size(hstate_vma(vma)));
 393
 394         if (!ptep)
 395                 return false;
 396
 397         pte = huge_ptep_get(ptep);
 398         if (huge_pte_none(pte) || !pte_present(pte))
 399                 return false;
 400
 401         if (pte_page(pte) == page)
 402                 return true;
 403
 404         return false;
 405 }
 406
 407 /*
 408  * Can vma_offset_start/vma_offset_end overflow on 32-bit arches?
 409  * No, because the interval tree returns us only those vmas
 410  * which overlap the truncated area starting at pgoff,
 411  * and no vma on a 32-bit arch can span beyond the 4GB.
 412  */
 413 static unsigned long vma_offset_start(struct vm_area_struct *vma, pgoff_t start)
 414 {
 415         if (vma->vm_pgoff < start)
 416                 return (start - vma->vm_pgoff) << PAGE_SHIFT;
 417         else
 418                 return 0;
 419 }
 420
 421 static unsigned long vma_offset_end(struct vm_area_struct *vma, pgoff_t end)
 422 {
 423         unsigned long t_end;
 424
 425         if (!end)
 426                 return vma->vm_end;
 427
 428         t_end = ((end - vma->vm_pgoff) << PAGE_SHIFT) + vma->vm_start;
 429         if (t_end > vma->vm_end)
 430                 t_end = vma->vm_end;
 431         return t_end;
 432 }
 433
 434 /*
 435  * Called with hugetlb fault mutex held.  Therefore, no more mappings to
 436  * this folio can be created while executing the routine.
 437  */
 438 static void hugetlb_unmap_file_folio(struct hstate *h,
 439                                         struct address_space *mapping,
 440                                         struct folio *folio, pgoff_t index)
 441 {
 442         struct rb_root_cached *root = &mapping->i_mmap;
 443         struct hugetlb_vma_lock *vma_lock;
 444         struct page *page = &folio->page;
 445         struct vm_area_struct *vma;
 446         unsigned long v_start;
 447         unsigned long v_end;
 448         pgoff_t start, end;
 449
 450         start = index * pages_per_huge_page(h);
 451         end = (index + 1) * pages_per_huge_page(h);
 452
 453         i_mmap_lock_write(mapping);
 454 retry:
 455         vma_lock = NULL;
 456         vma_interval_tree_foreach(vma, root, start, end - 1) {
 457                 v_start = vma_offset_start(vma, start);
 458                 v_end = vma_offset_end(vma, end);
 459
 460                 if (!hugetlb_vma_maps_page(vma, vma->vm_start + v_start, page))
 461                         continue;
 462
 463                 if (!hugetlb_vma_trylock_write(vma)) {
 464                         vma_lock = vma->vm_private_data;
 465                         /*
 466                          * If we can not get vma lock, we need to drop
 467                          * immap_sema and take locks in order.  First,
 468                          * take a ref on the vma_lock structure so that
 469                          * we can be guaranteed it will not go away when
 470                          * dropping immap_sema.
 471                          */
 472                         kref_get(&vma_lock->refs);
 473                         break;
 474                 }
 475
 476                 unmap_hugepage_range(vma, vma->vm_start + v_start, v_end,
 477                                 NULL, ZAP_FLAG_DROP_MARKER);
 478                 hugetlb_vma_unlock_write(vma);
 479         }
 480
 481         i_mmap_unlock_write(mapping);
 482
 483         if (vma_lock) {
 484                 /*
 485                  * Wait on vma_lock.  We know it is still valid as we have
 486                  * a reference.  We must 'open code' vma locking as we do
 487                  * not know if vma_lock is still attached to vma.
 488                  */
 489                 down_write(&vma_lock->rw_sema);
 490                 i_mmap_lock_write(mapping);
 491
 492                 vma = vma_lock->vma;
 493                 if (!vma) {
 494                         /*
 495                          * If lock is no longer attached to vma, then just
 496                          * unlock, drop our reference and retry looking for
 497                          * other vmas.
 498                          */
 499                         up_write(&vma_lock->rw_sema);
 500                         kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
 501                         goto retry;
 502                 }
 503
 504                 /*
 505                  * vma_lock is still attached to vma.  Check to see if vma
 506                  * still maps page and if so, unmap.
 507                  */
 508                 v_start = vma_offset_start(vma, start);
 509                 v_end = vma_offset_end(vma, end);
 510                 if (hugetlb_vma_maps_page(vma, vma->vm_start + v_start, page))
 511                         unmap_hugepage_range(vma, vma->vm_start + v_start,
 512                                                 v_end, NULL,
 513                                                 ZAP_FLAG_DROP_MARKER);
 514
 515                 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
 516                 hugetlb_vma_unlock_write(vma);
 517
 518                 goto retry;
 519         }
 520 }
 521
 522 static void
 523 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
 524                       zap_flags_t zap_flags)
 525 {
 526         struct vm_area_struct *vma;
 527
 528         /*
 529          * end == 0 indicates that the entire range after start should be
 530          * unmapped.  Note, end is exclusive, whereas the interval tree takes
 531          * an inclusive "last".
 532          */
 533         vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
 534                 unsigned long v_start;
 535                 unsigned long v_end;
 536
 537                 if (!hugetlb_vma_trylock_write(vma))
 538                         continue;
 539
 540                 v_start = vma_offset_start(vma, start);
 541                 v_end = vma_offset_end(vma, end);
 542
 543                 unmap_hugepage_range(vma, vma->vm_start + v_start, v_end,
 544                                      NULL, zap_flags);
 545
 546                 /*
 547                  * Note that vma lock only exists for shared/non-private
 548                  * vmas.  Therefore, lock is not held when calling
 549                  * unmap_hugepage_range for private vmas.
 550                  */
 551                 hugetlb_vma_unlock_write(vma);
 552         }
 553 }
 554
 555 /*
 556  * Called with hugetlb fault mutex held.
 557  * Returns true if page was actually removed, false otherwise.
 558  */
 559 static bool remove_inode_single_folio(struct hstate *h, struct inode *inode,
 560                                         struct address_space *mapping,
 561                                         struct folio *folio, pgoff_t index,
 562                                         bool truncate_op)
 563 {
 564         bool ret = false;
 565
 566         /*
 567          * If folio is mapped, it was faulted in after being
 568          * unmapped in caller.  Unmap (again) while holding
 569          * the fault mutex.  The mutex will prevent faults
 570          * until we finish removing the folio.
 571          */
 572         if (unlikely(folio_mapped(folio)))
 573                 hugetlb_unmap_file_folio(h, mapping, folio, index);
 574
 575         folio_lock(folio);
 576         /*
 577          * We must remove the folio from page cache before removing
 578          * the region/ reserve map (hugetlb_unreserve_pages).  In
 579          * rare out of memory conditions, removal of the region/reserve
 580          * map could fail.  Correspondingly, the subpool and global
 581          * reserve usage count can need to be adjusted.
 582          */
 583         VM_BUG_ON_FOLIO(folio_test_hugetlb_restore_reserve(folio), folio);
 584         hugetlb_delete_from_page_cache(folio);
 585         ret = true;
 586         if (!truncate_op) {
 587                 if (unlikely(hugetlb_unreserve_pages(inode, index,
 588                                                         index + 1, 1)))
 589                         hugetlb_fix_reserve_counts(inode);
 590         }
 591
 592         folio_unlock(folio);
 593         return ret;
 594 }
 595
 596 /*
 597  * remove_inode_hugepages handles two distinct cases: truncation and hole
 598  * punch.  There are subtle differences in operation for each case.
 599  *
 600  * truncation is indicated by end of range being LLONG_MAX
 601  *      In this case, we first scan the range and release found pages.
 602  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
 603  *      maps and global counts.  Page faults can race with truncation.
 604  *      During faults, hugetlb_no_page() checks i_size before page allocation,
 605  *      and again after obtaining page table lock.  It will 'back out'
 606  *      allocations in the truncated range.
 607  * hole punch is indicated if end is not LLONG_MAX
 608  *      In the hole punch case we scan the range and release found pages.
 609  *      Only when releasing a page is the associated region/reserve map
 610  *      deleted.  The region/reserve map for ranges without associated
 611  *      pages are not modified.  Page faults can race with hole punch.
 612  *      This is indicated if we find a mapped page.
 613  * Note: If the passed end of range value is beyond the end of file, but
 614  * not LLONG_MAX this routine still performs a hole punch operation.
 615  */
 616 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 617                                    loff_t lend)
 618 {
 619         struct hstate *h = hstate_inode(inode);
 620         struct address_space *mapping = &inode->i_data;
 621         const pgoff_t start = lstart >> huge_page_shift(h);
 622         const pgoff_t end = lend >> huge_page_shift(h);
 623         struct folio_batch fbatch;
 624         pgoff_t next, index;
 625         int i, freed = 0;
 626         bool truncate_op = (lend == LLONG_MAX);
 627
 628         folio_batch_init(&fbatch);
 629         next = start;
 630         while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
 631                 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
 632                         struct folio *folio = fbatch.folios[i];
 633                         u32 hash = 0;
 634
 635                         index = folio->index;
 636                         hash = hugetlb_fault_mutex_hash(mapping, index);
 637                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 638
 639                         /*
 640                          * Remove folio that was part of folio_batch.
 641                          */
 642                         if (remove_inode_single_folio(h, inode, mapping, folio,
 643                                                         index, truncate_op))
 644                                 freed++;
 645
 646                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 647                 }
 648                 folio_batch_release(&fbatch);
 649                 cond_resched();
 650         }
 651
 652         if (truncate_op)
 653                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 654 }
 655
 656 static void hugetlbfs_evict_inode(struct inode *inode)
 657 {
 658         struct resv_map *resv_map;
 659
 660         remove_inode_hugepages(inode, 0, LLONG_MAX);
 661
 662         /*
 663          * Get the resv_map from the address space embedded in the inode.
 664          * This is the address space which points to any resv_map allocated
 665          * at inode creation time.  If this is a device special inode,
 666          * i_mapping may not point to the original address space.
 667          */
 668         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
 669         /* Only regular and link inodes have associated reserve maps */
 670         if (resv_map)
 671                 resv_map_release(&resv_map->refs);
 672         clear_inode(inode);
 673 }
 674
 675 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 676 {
 677         pgoff_t pgoff;
 678         struct address_space *mapping = inode->i_mapping;
 679         struct hstate *h = hstate_inode(inode);
 680
 681         BUG_ON(offset & ~huge_page_mask(h));
 682         pgoff = offset >> PAGE_SHIFT;
 683
 684         i_size_write(inode, offset);
 685         i_mmap_lock_write(mapping);
 686         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 687                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
 688                                       ZAP_FLAG_DROP_MARKER);
 689         i_mmap_unlock_write(mapping);
 690         remove_inode_hugepages(inode, offset, LLONG_MAX);
 691 }
 692
 693 static void hugetlbfs_zero_partial_page(struct hstate *h,
 694                                         struct address_space *mapping,
 695                                         loff_t start,
 696                                         loff_t end)
 697 {
 698         pgoff_t idx = start >> huge_page_shift(h);
 699         struct folio *folio;
 700
 701         folio = filemap_lock_folio(mapping, idx);
 702         if (!folio)
 703                 return;
 704
 705         start = start & ~huge_page_mask(h);
 706         end = end & ~huge_page_mask(h);
 707         if (!end)
 708                 end = huge_page_size(h);
 709
 710         folio_zero_segment(folio, (size_t)start, (size_t)end);
 711
 712         folio_unlock(folio);
 713         folio_put(folio);
 714 }
 715
 716 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 717 {
 718         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 719         struct address_space *mapping = inode->i_mapping;
 720         struct hstate *h = hstate_inode(inode);
 721         loff_t hpage_size = huge_page_size(h);
 722         loff_t hole_start, hole_end;
 723
 724         /*
 725          * hole_start and hole_end indicate the full pages within the hole.
 726          */
 727         hole_start = round_up(offset, hpage_size);
 728         hole_end = round_down(offset + len, hpage_size);
 729
 730         inode_lock(inode);
 731
 732         /* protected by i_rwsem */
 733         if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 734                 inode_unlock(inode);
 735                 return -EPERM;
 736         }
 737
 738         i_mmap_lock_write(mapping);
 739
 740         /* If range starts before first full page, zero partial page. */
 741         if (offset < hole_start)
 742                 hugetlbfs_zero_partial_page(h, mapping,
 743                                 offset, min(offset + len, hole_start));
 744
 745         /* Unmap users of full pages in the hole. */
 746         if (hole_end > hole_start) {
 747                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 748                         hugetlb_vmdelete_list(&mapping->i_mmap,
 749                                               hole_start >> PAGE_SHIFT,
 750                                               hole_end >> PAGE_SHIFT, 0);
 751         }
 752
 753         /* If range extends beyond last full page, zero partial page. */
 754         if ((offset + len) > hole_end && (offset + len) > hole_start)
 755                 hugetlbfs_zero_partial_page(h, mapping,
 756                                 hole_end, offset + len);
 757
 758         i_mmap_unlock_write(mapping);
 759
 760         /* Remove full pages from the file. */
 761         if (hole_end > hole_start)
 762                 remove_inode_hugepages(inode, hole_start, hole_end);
 763
 764         inode_unlock(inode);
 765
 766         return 0;
 767 }
 768
 769 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 770                                 loff_t len)
 771 {
 772         struct inode *inode = file_inode(file);
 773         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 774         struct address_space *mapping = inode->i_mapping;
 775         struct hstate *h = hstate_inode(inode);
 776         struct vm_area_struct pseudo_vma;
 777         struct mm_struct *mm = current->mm;
 778         loff_t hpage_size = huge_page_size(h);
 779         unsigned long hpage_shift = huge_page_shift(h);
 780         pgoff_t start, index, end;
 781         int error;
 782         u32 hash;
 783
 784         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 785                 return -EOPNOTSUPP;
 786
 787         if (mode & FALLOC_FL_PUNCH_HOLE)
 788                 return hugetlbfs_punch_hole(inode, offset, len);
 789
 790         /*
 791          * Default preallocate case.
 792          * For this range, start is rounded down and end is rounded up
 793          * as well as being converted to page offsets.
 794          */
 795         start = offset >> hpage_shift;
 796         end = (offset + len + hpage_size - 1) >> hpage_shift;
 797
 798         inode_lock(inode);
 799
 800         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 801         error = inode_newsize_ok(inode, offset + len);
 802         if (error)
 803                 goto out;
 804
 805         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 806                 error = -EPERM;
 807                 goto out;
 808         }
 809
 810         /*
 811          * Initialize a pseudo vma as this is required by the huge page
 812          * allocation routines.  If NUMA is configured, use page index
 813          * as input to create an allocation policy.
 814          */
 815         vma_init(&pseudo_vma, mm);
 816         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 817         pseudo_vma.vm_file = file;
 818
 819         for (index = start; index < end; index++) {
 820                 /*
 821                  * This is supposed to be the vaddr where the page is being
 822                  * faulted in, but we have no vaddr here.
 823                  */
 824                 struct page *page;
 825                 unsigned long addr;
 826
 827                 cond_resched();
 828
 829                 /*
 830                  * fallocate(2) manpage permits EINTR; we may have been
 831                  * interrupted because we are using up too much memory.
 832                  */
 833                 if (signal_pending(current)) {
 834                         error = -EINTR;
 835                         break;
 836                 }
 837
 838                 /* Set numa allocation policy based on index */
 839                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 840
 841                 /* addr is the offset within the file (zero based) */
 842                 addr = index * hpage_size;
 843
 844                 /* mutex taken here, fault path and hole punch */
 845                 hash = hugetlb_fault_mutex_hash(mapping, index);
 846                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 847
 848                 /* See if already present in mapping to avoid alloc/free */
 849                 page = find_get_page(mapping, index);
 850                 if (page) {
 851                         put_page(page);
 852                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 853                         hugetlb_drop_vma_policy(&pseudo_vma);
 854                         continue;
 855                 }
 856
 857                 /*
 858                  * Allocate page without setting the avoid_reserve argument.
 859                  * There certainly are no reserves associated with the
 860                  * pseudo_vma.  However, there could be shared mappings with
 861                  * reserves for the file at the inode level.  If we fallocate
 862                  * pages in these areas, we need to consume the reserves
 863                  * to keep reservation accounting consistent.
 864                  */
 865                 page = alloc_huge_page(&pseudo_vma, addr, 0);
 866                 hugetlb_drop_vma_policy(&pseudo_vma);
 867                 if (IS_ERR(page)) {
 868                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 869                         error = PTR_ERR(page);
 870                         goto out;
 871                 }
 872                 clear_huge_page(page, addr, pages_per_huge_page(h));
 873                 __SetPageUptodate(page);
 874                 error = hugetlb_add_to_page_cache(page, mapping, index);
 875                 if (unlikely(error)) {
 876                         restore_reserve_on_error(h, &pseudo_vma, addr, page);
 877                         put_page(page);
 878                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 879                         goto out;
 880                 }
 881
 882                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 883
 884                 SetHPageMigratable(page);
 885                 /*
 886                  * unlock_page because locked by hugetlb_add_to_page_cache()
 887                  * put_page() due to reference from alloc_huge_page()
 888                  */
 889                 unlock_page(page);
 890                 put_page(page);
 891         }
 892
 893         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 894                 i_size_write(inode, offset + len);
 895         inode->i_ctime = current_time(inode);
 896 out:
 897         inode_unlock(inode);
 898         return error;
 899 }
 900
 901 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
 902                              struct dentry *dentry, struct iattr *attr)
 903 {
 904         struct inode *inode = d_inode(dentry);
 905         struct hstate *h = hstate_inode(inode);
 906         int error;
 907         unsigned int ia_valid = attr->ia_valid;
 908         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 909
 910         error = setattr_prepare(&init_user_ns, dentry, attr);
 911         if (error)
 912                 return error;
 913
 914         if (ia_valid & ATTR_SIZE) {
 915                 loff_t oldsize = inode->i_size;
 916                 loff_t newsize = attr->ia_size;
 917
 918                 if (newsize & ~huge_page_mask(h))
 919                         return -EINVAL;
 920                 /* protected by i_rwsem */
 921                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 922                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 923                         return -EPERM;
 924                 hugetlb_vmtruncate(inode, newsize);
 925         }
 926
 927         setattr_copy(&init_user_ns, inode, attr);
 928         mark_inode_dirty(inode);
 929         return 0;
 930 }
 931
 932 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 933                                         struct hugetlbfs_fs_context *ctx)
 934 {
 935         struct inode *inode;
 936
 937         inode = new_inode(sb);
 938         if (inode) {
 939                 inode->i_ino = get_next_ino();
 940                 inode->i_mode = S_IFDIR | ctx->mode;
 941                 inode->i_uid = ctx->uid;
 942                 inode->i_gid = ctx->gid;
 943                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 944                 inode->i_op = &hugetlbfs_dir_inode_operations;
 945                 inode->i_fop = &simple_dir_operations;
 946                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 947                 inc_nlink(inode);
 948                 lockdep_annotate_inode_mutex_key(inode);
 949         }
 950         return inode;
 951 }
 952
 953 /*
 954  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 955  * be taken from reclaim -- unlike regular filesystems. This needs an
 956  * annotation because huge_pmd_share() does an allocation under hugetlb's
 957  * i_mmap_rwsem.
 958  */
 959 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 960
 961 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 962                                         struct inode *dir,
 963                                         umode_t mode, dev_t dev)
 964 {
 965         struct inode *inode;
 966         struct resv_map *resv_map = NULL;
 967
 968         /*
 969          * Reserve maps are only needed for inodes that can have associated
 970          * page allocations.
 971          */
 972         if (S_ISREG(mode) || S_ISLNK(mode)) {
 973                 resv_map = resv_map_alloc();
 974                 if (!resv_map)
 975                         return NULL;
 976         }
 977
 978         inode = new_inode(sb);
 979         if (inode) {
 980                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 981
 982                 inode->i_ino = get_next_ino();
 983                 inode_init_owner(&init_user_ns, inode, dir, mode);
 984                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 985                                 &hugetlbfs_i_mmap_rwsem_key);
 986                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 987                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 988                 inode->i_mapping->private_data = resv_map;
 989                 info->seals = F_SEAL_SEAL;
 990                 switch (mode & S_IFMT) {
 991                 default:
 992                         init_special_inode(inode, mode, dev);
 993                         break;
 994                 case S_IFREG:
 995                         inode->i_op = &hugetlbfs_inode_operations;
 996                         inode->i_fop = &hugetlbfs_file_operations;
 997                         break;
 998                 case S_IFDIR:
 999                         inode->i_op = &hugetlbfs_dir_inode_operations;
1000                         inode->i_fop = &simple_dir_operations;
1001
1002                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
1003                         inc_nlink(inode);
1004                         break;
1005                 case S_IFLNK:
1006                         inode->i_op = &page_symlink_inode_operations;
1007                         inode_nohighmem(inode);
1008                         break;
1009                 }
1010                 lockdep_annotate_inode_mutex_key(inode);
1011         } else {
1012                 if (resv_map)
1013                         kref_put(&resv_map->refs, resv_map_release);
1014         }
1015
1016         return inode;
1017 }
1018
1019 /*
1020  * File creation. Allocate an inode, and we're done..
1021  */
1022 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
1023                            struct dentry *dentry, umode_t mode, dev_t dev)
1024 {
1025         struct inode *inode;
1026
1027         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
1028         if (!inode)
1029                 return -ENOSPC;
1030         dir->i_ctime = dir->i_mtime = current_time(dir);
1031         d_instantiate(dentry, inode);
1032         dget(dentry);/* Extra count - pin the dentry in core */
1033         return 0;
1034 }
1035
1036 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
1037                            struct dentry *dentry, umode_t mode)
1038 {
1039         int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
1040                                      mode | S_IFDIR, 0);
1041         if (!retval)
1042                 inc_nlink(dir);
1043         return retval;
1044 }
1045
1046 static int hugetlbfs_create(struct user_namespace *mnt_userns,
1047                             struct inode *dir, struct dentry *dentry,
1048                             umode_t mode, bool excl)
1049 {
1050         return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
1051 }
1052
1053 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
1054                              struct inode *dir, struct file *file,
1055                              umode_t mode)
1056 {
1057         struct inode *inode;
1058
1059         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode | S_IFREG, 0);
1060         if (!inode)
1061                 return -ENOSPC;
1062         dir->i_ctime = dir->i_mtime = current_time(dir);
1063         d_tmpfile(file, inode);
1064         return finish_open_simple(file, 0);
1065 }
1066
1067 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
1068                              struct inode *dir, struct dentry *dentry,
1069                              const char *symname)
1070 {
1071         struct inode *inode;
1072         int error = -ENOSPC;
1073
1074         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
1075         if (inode) {
1076                 int l = strlen(symname)+1;
1077                 error = page_symlink(inode, symname, l);
1078                 if (!error) {
1079                         d_instantiate(dentry, inode);
1080                         dget(dentry);
1081                 } else
1082                         iput(inode);
1083         }
1084         dir->i_ctime = dir->i_mtime = current_time(dir);
1085
1086         return error;
1087 }
1088
1089 #ifdef CONFIG_MIGRATION
1090 static int hugetlbfs_migrate_folio(struct address_space *mapping,
1091                                 struct folio *dst, struct folio *src,
1092                                 enum migrate_mode mode)
1093 {
1094         int rc;
1095
1096         rc = migrate_huge_page_move_mapping(mapping, dst, src);
1097         if (rc != MIGRATEPAGE_SUCCESS)
1098                 return rc;
1099
1100         if (hugetlb_folio_subpool(src)) {
1101                 hugetlb_set_folio_subpool(dst,
1102                                         hugetlb_folio_subpool(src));
1103                 hugetlb_set_folio_subpool(src, NULL);
1104         }
1105
1106         if (mode != MIGRATE_SYNC_NO_COPY)
1107                 folio_migrate_copy(dst, src);
1108         else
1109                 folio_migrate_flags(dst, src);
1110
1111         return MIGRATEPAGE_SUCCESS;
1112 }
1113 #else
1114 #define hugetlbfs_migrate_folio NULL
1115 #endif
1116
1117 static int hugetlbfs_error_remove_page(struct address_space *mapping,
1118                                 struct page *page)
1119 {
1120         return 0;
1121 }
1122
1123 /*
1124  * Display the mount options in /proc/mounts.
1125  */
1126 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1127 {
1128         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1129         struct hugepage_subpool *spool = sbinfo->spool;
1130         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1131         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1132         char mod;
1133
1134         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1135                 seq_printf(m, ",uid=%u",
1136                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1137         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1138                 seq_printf(m, ",gid=%u",
1139                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1140         if (sbinfo->mode != 0755)
1141                 seq_printf(m, ",mode=%o", sbinfo->mode);
1142         if (sbinfo->max_inodes != -1)
1143                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1144
1145         hpage_size /= 1024;
1146         mod = 'K';
1147         if (hpage_size >= 1024) {
1148                 hpage_size /= 1024;
1149                 mod = 'M';
1150         }
1151         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1152         if (spool) {
1153                 if (spool->max_hpages != -1)
1154                         seq_printf(m, ",size=%llu",
1155                                    (unsigned long long)spool->max_hpages << hpage_shift);
1156                 if (spool->min_hpages != -1)
1157                         seq_printf(m, ",min_size=%llu",
1158                                    (unsigned long long)spool->min_hpages << hpage_shift);
1159         }
1160         return 0;
1161 }
1162
1163 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1164 {
1165         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1166         struct hstate *h = hstate_inode(d_inode(dentry));
1167
1168         buf->f_type = HUGETLBFS_MAGIC;
1169         buf->f_bsize = huge_page_size(h);
1170         if (sbinfo) {
1171                 spin_lock(&sbinfo->stat_lock);
1172                 /* If no limits set, just report 0 or -1 for max/free/used
1173                  * blocks, like simple_statfs() */
1174                 if (sbinfo->spool) {
1175                         long free_pages;
1176
1177                         spin_lock_irq(&sbinfo->spool->lock);
1178                         buf->f_blocks = sbinfo->spool->max_hpages;
1179                         free_pages = sbinfo->spool->max_hpages
1180                                 - sbinfo->spool->used_hpages;
1181                         buf->f_bavail = buf->f_bfree = free_pages;
1182                         spin_unlock_irq(&sbinfo->spool->lock);
1183                         buf->f_files = sbinfo->max_inodes;
1184                         buf->f_ffree = sbinfo->free_inodes;
1185                 }
1186                 spin_unlock(&sbinfo->stat_lock);
1187         }
1188         buf->f_namelen = NAME_MAX;
1189         return 0;
1190 }
1191
1192 static void hugetlbfs_put_super(struct super_block *sb)
1193 {
1194         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1195
1196         if (sbi) {
1197                 sb->s_fs_info = NULL;
1198
1199                 if (sbi->spool)
1200                         hugepage_put_subpool(sbi->spool);
1201
1202                 kfree(sbi);
1203         }
1204 }
1205
1206 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1207 {
1208         if (sbinfo->free_inodes >= 0) {
1209                 spin_lock(&sbinfo->stat_lock);
1210                 if (unlikely(!sbinfo->free_inodes)) {
1211                         spin_unlock(&sbinfo->stat_lock);
1212                         return 0;
1213                 }
1214                 sbinfo->free_inodes--;
1215                 spin_unlock(&sbinfo->stat_lock);
1216         }
1217
1218         return 1;
1219 }
1220
1221 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1222 {
1223         if (sbinfo->free_inodes >= 0) {
1224                 spin_lock(&sbinfo->stat_lock);
1225                 sbinfo->free_inodes++;
1226                 spin_unlock(&sbinfo->stat_lock);
1227         }
1228 }
1229
1230
1231 static struct kmem_cache *hugetlbfs_inode_cachep;
1232
1233 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1234 {
1235         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1236         struct hugetlbfs_inode_info *p;
1237
1238         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1239                 return NULL;
1240         p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1241         if (unlikely(!p)) {
1242                 hugetlbfs_inc_free_inodes(sbinfo);
1243                 return NULL;
1244         }
1245
1246         /*
1247          * Any time after allocation, hugetlbfs_destroy_inode can be called
1248          * for the inode.  mpol_free_shared_policy is unconditionally called
1249          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1250          * in case of a quick call to destroy.
1251          *
1252          * Note that the policy is initialized even if we are creating a
1253          * private inode.  This simplifies hugetlbfs_destroy_inode.
1254          */
1255         mpol_shared_policy_init(&p->policy, NULL);
1256
1257         return &p->vfs_inode;
1258 }
1259
1260 static void hugetlbfs_free_inode(struct inode *inode)
1261 {
1262         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1263 }
1264
1265 static void hugetlbfs_destroy_inode(struct inode *inode)
1266 {
1267         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1268         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1269 }
1270
1271 static const struct address_space_operations hugetlbfs_aops = {
1272         .write_begin    = hugetlbfs_write_begin,
1273         .write_end      = hugetlbfs_write_end,
1274         .dirty_folio    = noop_dirty_folio,
1275         .migrate_folio  = hugetlbfs_migrate_folio,
1276         .error_remove_page      = hugetlbfs_error_remove_page,
1277 };
1278
1279
1280 static void init_once(void *foo)
1281 {
1282         struct hugetlbfs_inode_info *ei = foo;
1283
1284         inode_init_once(&ei->vfs_inode);
1285 }
1286
1287 const struct file_operations hugetlbfs_file_operations = {
1288         .read_iter              = hugetlbfs_read_iter,
1289         .mmap                   = hugetlbfs_file_mmap,
1290         .fsync                  = noop_fsync,
1291         .get_unmapped_area      = hugetlb_get_unmapped_area,
1292         .llseek                 = default_llseek,
1293         .fallocate              = hugetlbfs_fallocate,
1294 };
1295
1296 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1297         .create         = hugetlbfs_create,
1298         .lookup         = simple_lookup,
1299         .link           = simple_link,
1300         .unlink         = simple_unlink,
1301         .symlink        = hugetlbfs_symlink,
1302         .mkdir          = hugetlbfs_mkdir,
1303         .rmdir          = simple_rmdir,
1304         .mknod          = hugetlbfs_mknod,
1305         .rename         = simple_rename,
1306         .setattr        = hugetlbfs_setattr,
1307         .tmpfile        = hugetlbfs_tmpfile,
1308 };
1309
1310 static const struct inode_operations hugetlbfs_inode_operations = {
1311         .setattr        = hugetlbfs_setattr,
1312 };
1313
1314 static const struct super_operations hugetlbfs_ops = {
1315         .alloc_inode    = hugetlbfs_alloc_inode,
1316         .free_inode     = hugetlbfs_free_inode,
1317         .destroy_inode  = hugetlbfs_destroy_inode,
1318         .evict_inode    = hugetlbfs_evict_inode,
1319         .statfs         = hugetlbfs_statfs,
1320         .put_super      = hugetlbfs_put_super,
1321         .show_options   = hugetlbfs_show_options,
1322 };
1323
1324 /*
1325  * Convert size option passed from command line to number of huge pages
1326  * in the pool specified by hstate.  Size option could be in bytes
1327  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1328  */
1329 static long
1330 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1331                          enum hugetlbfs_size_type val_type)
1332 {
1333         if (val_type == NO_SIZE)
1334                 return -1;
1335
1336         if (val_type == SIZE_PERCENT) {
1337                 size_opt <<= huge_page_shift(h);
1338                 size_opt *= h->max_huge_pages;
1339                 do_div(size_opt, 100);
1340         }
1341
1342         size_opt >>= huge_page_shift(h);
1343         return size_opt;
1344 }
1345
1346 /*
1347  * Parse one mount parameter.
1348  */
1349 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1350 {
1351         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1352         struct fs_parse_result result;
1353         char *rest;
1354         unsigned long ps;
1355         int opt;
1356
1357         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1358         if (opt < 0)
1359                 return opt;
1360
1361         switch (opt) {
1362         case Opt_uid:
1363                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1364                 if (!uid_valid(ctx->uid))
1365                         goto bad_val;
1366                 return 0;
1367
1368         case Opt_gid:
1369                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1370                 if (!gid_valid(ctx->gid))
1371                         goto bad_val;
1372                 return 0;
1373
1374         case Opt_mode:
1375                 ctx->mode = result.uint_32 & 01777U;
1376                 return 0;
1377
1378         case Opt_size:
1379                 /* memparse() will accept a K/M/G without a digit */
1380                 if (!param->string || !isdigit(param->string[0]))
1381                         goto bad_val;
1382                 ctx->max_size_opt = memparse(param->string, &rest);
1383                 ctx->max_val_type = SIZE_STD;
1384                 if (*rest == '%')
1385                         ctx->max_val_type = SIZE_PERCENT;
1386                 return 0;
1387
1388         case Opt_nr_inodes:
1389                 /* memparse() will accept a K/M/G without a digit */
1390                 if (!param->string || !isdigit(param->string[0]))
1391                         goto bad_val;
1392                 ctx->nr_inodes = memparse(param->string, &rest);
1393                 return 0;
1394
1395         case Opt_pagesize:
1396                 ps = memparse(param->string, &rest);
1397                 ctx->hstate = size_to_hstate(ps);
1398                 if (!ctx->hstate) {
1399                         pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1400                         return -EINVAL;
1401                 }
1402                 return 0;
1403
1404         case Opt_min_size:
1405                 /* memparse() will accept a K/M/G without a digit */
1406                 if (!param->string || !isdigit(param->string[0]))
1407                         goto bad_val;
1408                 ctx->min_size_opt = memparse(param->string, &rest);
1409                 ctx->min_val_type = SIZE_STD;
1410                 if (*rest == '%')
1411                         ctx->min_val_type = SIZE_PERCENT;
1412                 return 0;
1413
1414         default:
1415                 return -EINVAL;
1416         }
1417
1418 bad_val:
1419         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1420                       param->string, param->key);
1421 }
1422
1423 /*
1424  * Validate the parsed options.
1425  */
1426 static int hugetlbfs_validate(struct fs_context *fc)
1427 {
1428         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1429
1430         /*
1431          * Use huge page pool size (in hstate) to convert the size
1432          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1433          */
1434         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1435                                                    ctx->max_size_opt,
1436                                                    ctx->max_val_type);
1437         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1438                                                    ctx->min_size_opt,
1439                                                    ctx->min_val_type);
1440
1441         /*
1442          * If max_size was specified, then min_size must be smaller
1443          */
1444         if (ctx->max_val_type > NO_SIZE &&
1445             ctx->min_hpages > ctx->max_hpages) {
1446                 pr_err("Minimum size can not be greater than maximum size\n");
1447                 return -EINVAL;
1448         }
1449
1450         return 0;
1451 }
1452
1453 static int
1454 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1455 {
1456         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1457         struct hugetlbfs_sb_info *sbinfo;
1458
1459         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1460         if (!sbinfo)
1461                 return -ENOMEM;
1462         sb->s_fs_info = sbinfo;
1463         spin_lock_init(&sbinfo->stat_lock);
1464         sbinfo->hstate          = ctx->hstate;
1465         sbinfo->max_inodes      = ctx->nr_inodes;
1466         sbinfo->free_inodes     = ctx->nr_inodes;
1467         sbinfo->spool           = NULL;
1468         sbinfo->uid             = ctx->uid;
1469         sbinfo->gid             = ctx->gid;
1470         sbinfo->mode            = ctx->mode;
1471
1472         /*
1473          * Allocate and initialize subpool if maximum or minimum size is
1474          * specified.  Any needed reservations (for minimum size) are taken
1475          * when the subpool is created.
1476          */
1477         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1478                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1479                                                      ctx->max_hpages,
1480                                                      ctx->min_hpages);
1481                 if (!sbinfo->spool)
1482                         goto out_free;
1483         }
1484         sb->s_maxbytes = MAX_LFS_FILESIZE;
1485         sb->s_blocksize = huge_page_size(ctx->hstate);
1486         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1487         sb->s_magic = HUGETLBFS_MAGIC;
1488         sb->s_op = &hugetlbfs_ops;
1489         sb->s_time_gran = 1;
1490
1491         /*
1492          * Due to the special and limited functionality of hugetlbfs, it does
1493          * not work well as a stacking filesystem.
1494          */
1495         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1496         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1497         if (!sb->s_root)
1498                 goto out_free;
1499         return 0;
1500 out_free:
1501         kfree(sbinfo->spool);
1502         kfree(sbinfo);
1503         return -ENOMEM;
1504 }
1505
1506 static int hugetlbfs_get_tree(struct fs_context *fc)
1507 {
1508         int err = hugetlbfs_validate(fc);
1509         if (err)
1510                 return err;
1511         return get_tree_nodev(fc, hugetlbfs_fill_super);
1512 }
1513
1514 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1515 {
1516         kfree(fc->fs_private);
1517 }
1518
1519 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1520         .free           = hugetlbfs_fs_context_free,
1521         .parse_param    = hugetlbfs_parse_param,
1522         .get_tree       = hugetlbfs_get_tree,
1523 };
1524
1525 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1526 {
1527         struct hugetlbfs_fs_context *ctx;
1528
1529         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1530         if (!ctx)
1531                 return -ENOMEM;
1532
1533         ctx->max_hpages = -1; /* No limit on size by default */
1534         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1535         ctx->uid        = current_fsuid();
1536         ctx->gid        = current_fsgid();
1537         ctx->mode       = 0755;
1538         ctx->hstate     = &default_hstate;
1539         ctx->min_hpages = -1; /* No default minimum size */
1540         ctx->max_val_type = NO_SIZE;
1541         ctx->min_val_type = NO_SIZE;
1542         fc->fs_private = ctx;
1543         fc->ops = &hugetlbfs_fs_context_ops;
1544         return 0;
1545 }
1546
1547 static struct file_system_type hugetlbfs_fs_type = {
1548         .name                   = "hugetlbfs",
1549         .init_fs_context        = hugetlbfs_init_fs_context,
1550         .parameters             = hugetlb_fs_parameters,
1551         .kill_sb                = kill_litter_super,
1552 };
1553
1554 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1555
1556 static int can_do_hugetlb_shm(void)
1557 {
1558         kgid_t shm_group;
1559         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1560         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1561 }
1562
1563 static int get_hstate_idx(int page_size_log)
1564 {
1565         struct hstate *h = hstate_sizelog(page_size_log);
1566
1567         if (!h)
1568                 return -1;
1569         return hstate_index(h);
1570 }
1571
1572 /*
1573  * Note that size should be aligned to proper hugepage size in caller side,
1574  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1575  */
1576 struct file *hugetlb_file_setup(const char *name, size_t size,
1577                                 vm_flags_t acctflag, int creat_flags,
1578                                 int page_size_log)
1579 {
1580         struct inode *inode;
1581         struct vfsmount *mnt;
1582         int hstate_idx;
1583         struct file *file;
1584
1585         hstate_idx = get_hstate_idx(page_size_log);
1586         if (hstate_idx < 0)
1587                 return ERR_PTR(-ENODEV);
1588
1589         mnt = hugetlbfs_vfsmount[hstate_idx];
1590         if (!mnt)
1591                 return ERR_PTR(-ENOENT);
1592
1593         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1594                 struct ucounts *ucounts = current_ucounts();
1595
1596                 if (user_shm_lock(size, ucounts)) {
1597                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1598                                 current->comm, current->pid);
1599                         user_shm_unlock(size, ucounts);
1600                 }
1601                 return ERR_PTR(-EPERM);
1602         }
1603
1604         file = ERR_PTR(-ENOSPC);
1605         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1606         if (!inode)
1607                 goto out;
1608         if (creat_flags == HUGETLB_SHMFS_INODE)
1609                 inode->i_flags |= S_PRIVATE;
1610
1611         inode->i_size = size;
1612         clear_nlink(inode);
1613
1614         if (!hugetlb_reserve_pages(inode, 0,
1615                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1616                         acctflag))
1617                 file = ERR_PTR(-ENOMEM);
1618         else
1619                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1620                                         &hugetlbfs_file_operations);
1621         if (!IS_ERR(file))
1622                 return file;
1623
1624         iput(inode);
1625 out:
1626         return file;
1627 }
1628
1629 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1630 {
1631         struct fs_context *fc;
1632         struct vfsmount *mnt;
1633
1634         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1635         if (IS_ERR(fc)) {
1636                 mnt = ERR_CAST(fc);
1637         } else {
1638                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1639                 ctx->hstate = h;
1640                 mnt = fc_mount(fc);
1641                 put_fs_context(fc);
1642         }
1643         if (IS_ERR(mnt))
1644                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1645                        huge_page_size(h) / SZ_1K);
1646         return mnt;
1647 }
1648
1649 static int __init init_hugetlbfs_fs(void)
1650 {
1651         struct vfsmount *mnt;
1652         struct hstate *h;
1653         int error;
1654         int i;
1655
1656         if (!hugepages_supported()) {
1657                 pr_info("disabling because there are no supported hugepage sizes\n");
1658                 return -ENOTSUPP;
1659         }
1660
1661         error = -ENOMEM;
1662         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1663                                         sizeof(struct hugetlbfs_inode_info),
1664                                         0, SLAB_ACCOUNT, init_once);
1665         if (hugetlbfs_inode_cachep == NULL)
1666                 goto out;
1667
1668         error = register_filesystem(&hugetlbfs_fs_type);
1669         if (error)
1670                 goto out_free;
1671
1672         /* default hstate mount is required */
1673         mnt = mount_one_hugetlbfs(&default_hstate);
1674         if (IS_ERR(mnt)) {
1675                 error = PTR_ERR(mnt);
1676                 goto out_unreg;
1677         }
1678         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1679
1680         /* other hstates are optional */
1681         i = 0;
1682         for_each_hstate(h) {
1683                 if (i == default_hstate_idx) {
1684                         i++;
1685                         continue;
1686                 }
1687
1688                 mnt = mount_one_hugetlbfs(h);
1689                 if (IS_ERR(mnt))
1690                         hugetlbfs_vfsmount[i] = NULL;
1691                 else
1692                         hugetlbfs_vfsmount[i] = mnt;
1693                 i++;
1694         }
1695
1696         return 0;
1697
1698  out_unreg:
1699         (void)unregister_filesystem(&hugetlbfs_fs_type);
1700  out_free:
1701         kmem_cache_destroy(hugetlbfs_inode_cachep);
1702  out:
1703         return error;
1704 }
1705 fs_initcall(init_hugetlbfs_fs)