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 static 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 /*
  87  * Mask used when checking the page offset value passed in via system
  88  * calls.  This value will be converted to a loff_t which is signed.
  89  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
  90  * value.  The extra bit (- 1 in the shift value) is to take the sign
  91  * bit into account.
  92  */
  93 #define PGOFF_LOFFT_MAX \
  94         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
  95
  96 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
  97 {
  98         struct inode *inode = file_inode(file);
  99         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 100         loff_t len, vma_len;
 101         int ret;
 102         struct hstate *h = hstate_file(file);
 103         vm_flags_t vm_flags;
 104
 105         /*
 106          * vma address alignment (but not the pgoff alignment) has
 107          * already been checked by prepare_hugepage_range.  If you add
 108          * any error returns here, do so after setting VM_HUGETLB, so
 109          * is_vm_hugetlb_page tests below unmap_region go the right
 110          * way when do_mmap unwinds (may be important on powerpc
 111          * and ia64).
 112          */
 113         vm_flags_set(vma, VM_HUGETLB | VM_DONTEXPAND);
 114         vma->vm_ops = &hugetlb_vm_ops;
 115
 116         ret = seal_check_write(info->seals, vma);
 117         if (ret)
 118                 return ret;
 119
 120         /*
 121          * page based offset in vm_pgoff could be sufficiently large to
 122          * overflow a loff_t when converted to byte offset.  This can
 123          * only happen on architectures where sizeof(loff_t) ==
 124          * sizeof(unsigned long).  So, only check in those instances.
 125          */
 126         if (sizeof(unsigned long) == sizeof(loff_t)) {
 127                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 128                         return -EINVAL;
 129         }
 130
 131         /* must be huge page aligned */
 132         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 133                 return -EINVAL;
 134
 135         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 136         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 137         /* check for overflow */
 138         if (len < vma_len)
 139                 return -EINVAL;
 140
 141         inode_lock(inode);
 142         file_accessed(file);
 143
 144         ret = -ENOMEM;
 145
 146         vm_flags = vma->vm_flags;
 147         /*
 148          * for SHM_HUGETLB, the pages are reserved in the shmget() call so skip
 149          * reserving here. Note: only for SHM hugetlbfs file, the inode
 150          * flag S_PRIVATE is set.
 151          */
 152         if (inode->i_flags & S_PRIVATE)
 153                 vm_flags |= VM_NORESERVE;
 154
 155         if (!hugetlb_reserve_pages(inode,
 156                                 vma->vm_pgoff >> huge_page_order(h),
 157                                 len >> huge_page_shift(h), vma,
 158                                 vm_flags))
 159                 goto out;
 160
 161         ret = 0;
 162         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 163                 i_size_write(inode, len);
 164 out:
 165         inode_unlock(inode);
 166
 167         return ret;
 168 }
 169
 170 /*
 171  * Called under mmap_write_lock(mm).
 172  */
 173
 174 static unsigned long
 175 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
 176                 unsigned long len, unsigned long pgoff, unsigned long flags)
 177 {
 178         struct hstate *h = hstate_file(file);
 179         struct vm_unmapped_area_info info = {};
 180
 181         info.length = len;
 182         info.low_limit = current->mm->mmap_base;
 183         info.high_limit = arch_get_mmap_end(addr, len, flags);
 184         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 185         return vm_unmapped_area(&info);
 186 }
 187
 188 static unsigned long
 189 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
 190                 unsigned long len, unsigned long pgoff, unsigned long flags)
 191 {
 192         struct hstate *h = hstate_file(file);
 193         struct vm_unmapped_area_info info = {};
 194
 195         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 196         info.length = len;
 197         info.low_limit = PAGE_SIZE;
 198         info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
 199         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 200         addr = vm_unmapped_area(&info);
 201
 202         /*
 203          * A failed mmap() very likely causes application failure,
 204          * so fall back to the bottom-up function here. This scenario
 205          * can happen with large stack limits and large mmap()
 206          * allocations.
 207          */
 208         if (unlikely(offset_in_page(addr))) {
 209                 VM_BUG_ON(addr != -ENOMEM);
 210                 info.flags = 0;
 211                 info.low_limit = current->mm->mmap_base;
 212                 info.high_limit = arch_get_mmap_end(addr, len, flags);
 213                 addr = vm_unmapped_area(&info);
 214         }
 215
 216         return addr;
 217 }
 218
 219 unsigned long
 220 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 221                                   unsigned long len, unsigned long pgoff,
 222                                   unsigned long flags)
 223 {
 224         struct mm_struct *mm = current->mm;
 225         struct vm_area_struct *vma;
 226         struct hstate *h = hstate_file(file);
 227         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
 228
 229         if (len & ~huge_page_mask(h))
 230                 return -EINVAL;
 231         if (len > TASK_SIZE)
 232                 return -ENOMEM;
 233
 234         if (flags & MAP_FIXED) {
 235                 if (prepare_hugepage_range(file, addr, len))
 236                         return -EINVAL;
 237                 return addr;
 238         }
 239
 240         if (addr) {
 241                 addr = ALIGN(addr, huge_page_size(h));
 242                 vma = find_vma(mm, addr);
 243                 if (mmap_end - len >= addr &&
 244                     (!vma || addr + len <= vm_start_gap(vma)))
 245                         return addr;
 246         }
 247
 248         /*
 249          * Use MMF_TOPDOWN flag as a hint to use topdown routine.
 250          * If architectures have special needs, they should define their own
 251          * version of hugetlb_get_unmapped_area.
 252          */
 253         if (test_bit(MMF_TOPDOWN, &mm->flags))
 254                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 255                                 pgoff, flags);
 256         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 257                         pgoff, flags);
 258 }
 259
 260 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 261 static unsigned long
 262 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 263                           unsigned long len, unsigned long pgoff,
 264                           unsigned long flags)
 265 {
 266         return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
 267 }
 268 #endif
 269
 270 /*
 271  * Someone wants to read @bytes from a HWPOISON hugetlb @page from @offset.
 272  * Returns the maximum number of bytes one can read without touching the 1st raw
 273  * HWPOISON subpage.
 274  *
 275  * The implementation borrows the iteration logic from copy_page_to_iter*.
 276  */
 277 static size_t adjust_range_hwpoison(struct page *page, size_t offset, size_t bytes)
 278 {
 279         size_t n = 0;
 280         size_t res = 0;
 281
 282         /* First subpage to start the loop. */
 283         page = nth_page(page, offset / PAGE_SIZE);
 284         offset %= PAGE_SIZE;
 285         while (1) {
 286                 if (is_raw_hwpoison_page_in_hugepage(page))
 287                         break;
 288
 289                 /* Safe to read n bytes without touching HWPOISON subpage. */
 290                 n = min(bytes, (size_t)PAGE_SIZE - offset);
 291                 res += n;
 292                 bytes -= n;
 293                 if (!bytes || !n)
 294                         break;
 295                 offset += n;
 296                 if (offset == PAGE_SIZE) {
 297                         page = nth_page(page, 1);
 298                         offset = 0;
 299                 }
 300         }
 301
 302         return res;
 303 }
 304
 305 /*
 306  * Support for read() - Find the page attached to f_mapping and copy out the
 307  * data. This provides functionality similar to filemap_read().
 308  */
 309 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 310 {
 311         struct file *file = iocb->ki_filp;
 312         struct hstate *h = hstate_file(file);
 313         struct address_space *mapping = file->f_mapping;
 314         struct inode *inode = mapping->host;
 315         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 316         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 317         unsigned long end_index;
 318         loff_t isize;
 319         ssize_t retval = 0;
 320
 321         while (iov_iter_count(to)) {
 322                 struct folio *folio;
 323                 size_t nr, copied, want;
 324
 325                 /* nr is the maximum number of bytes to copy from this page */
 326                 nr = huge_page_size(h);
 327                 isize = i_size_read(inode);
 328                 if (!isize)
 329                         break;
 330                 end_index = (isize - 1) >> huge_page_shift(h);
 331                 if (index > end_index)
 332                         break;
 333                 if (index == end_index) {
 334                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 335                         if (nr <= offset)
 336                                 break;
 337                 }
 338                 nr = nr - offset;
 339
 340                 /* Find the folio */
 341                 folio = filemap_lock_hugetlb_folio(h, mapping, index);
 342                 if (IS_ERR(folio)) {
 343                         /*
 344                          * We have a HOLE, zero out the user-buffer for the
 345                          * length of the hole or request.
 346                          */
 347                         copied = iov_iter_zero(nr, to);
 348                 } else {
 349                         folio_unlock(folio);
 350
 351                         if (!folio_test_hwpoison(folio))
 352                                 want = nr;
 353                         else {
 354                                 /*
 355                                  * Adjust how many bytes safe to read without
 356                                  * touching the 1st raw HWPOISON subpage after
 357                                  * offset.
 358                                  */
 359                                 want = adjust_range_hwpoison(&folio->page, offset, nr);
 360                                 if (want == 0) {
 361                                         folio_put(folio);
 362                                         retval = -EIO;
 363                                         break;
 364                                 }
 365                         }
 366
 367                         /*
 368                          * We have the folio, copy it to user space buffer.
 369                          */
 370                         copied = copy_folio_to_iter(folio, offset, want, to);
 371                         folio_put(folio);
 372                 }
 373                 offset += copied;
 374                 retval += copied;
 375                 if (copied != nr && iov_iter_count(to)) {
 376                         if (!retval)
 377                                 retval = -EFAULT;
 378                         break;
 379                 }
 380                 index += offset >> huge_page_shift(h);
 381                 offset &= ~huge_page_mask(h);
 382         }
 383         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 384         return retval;
 385 }
 386
 387 static int hugetlbfs_write_begin(struct file *file,
 388                         struct address_space *mapping,
 389                         loff_t pos, unsigned len,
 390                         struct page **pagep, void **fsdata)
 391 {
 392         return -EINVAL;
 393 }
 394
 395 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 396                         loff_t pos, unsigned len, unsigned copied,
 397                         struct page *page, void *fsdata)
 398 {
 399         BUG();
 400         return -EINVAL;
 401 }
 402
 403 static void hugetlb_delete_from_page_cache(struct folio *folio)
 404 {
 405         folio_clear_dirty(folio);
 406         folio_clear_uptodate(folio);
 407         filemap_remove_folio(folio);
 408 }
 409
 410 /*
 411  * Called with i_mmap_rwsem held for inode based vma maps.  This makes
 412  * sure vma (and vm_mm) will not go away.  We also hold the hugetlb fault
 413  * mutex for the page in the mapping.  So, we can not race with page being
 414  * faulted into the vma.
 415  */
 416 static bool hugetlb_vma_maps_page(struct vm_area_struct *vma,
 417                                 unsigned long addr, struct page *page)
 418 {
 419         pte_t *ptep, pte;
 420
 421         ptep = hugetlb_walk(vma, addr, huge_page_size(hstate_vma(vma)));
 422         if (!ptep)
 423                 return false;
 424
 425         pte = huge_ptep_get(ptep);
 426         if (huge_pte_none(pte) || !pte_present(pte))
 427                 return false;
 428
 429         if (pte_page(pte) == page)
 430                 return true;
 431
 432         return false;
 433 }
 434
 435 /*
 436  * Can vma_offset_start/vma_offset_end overflow on 32-bit arches?
 437  * No, because the interval tree returns us only those vmas
 438  * which overlap the truncated area starting at pgoff,
 439  * and no vma on a 32-bit arch can span beyond the 4GB.
 440  */
 441 static unsigned long vma_offset_start(struct vm_area_struct *vma, pgoff_t start)
 442 {
 443         unsigned long offset = 0;
 444
 445         if (vma->vm_pgoff < start)
 446                 offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 447
 448         return vma->vm_start + offset;
 449 }
 450
 451 static unsigned long vma_offset_end(struct vm_area_struct *vma, pgoff_t end)
 452 {
 453         unsigned long t_end;
 454
 455         if (!end)
 456                 return vma->vm_end;
 457
 458         t_end = ((end - vma->vm_pgoff) << PAGE_SHIFT) + vma->vm_start;
 459         if (t_end > vma->vm_end)
 460                 t_end = vma->vm_end;
 461         return t_end;
 462 }
 463
 464 /*
 465  * Called with hugetlb fault mutex held.  Therefore, no more mappings to
 466  * this folio can be created while executing the routine.
 467  */
 468 static void hugetlb_unmap_file_folio(struct hstate *h,
 469                                         struct address_space *mapping,
 470                                         struct folio *folio, pgoff_t index)
 471 {
 472         struct rb_root_cached *root = &mapping->i_mmap;
 473         struct hugetlb_vma_lock *vma_lock;
 474         struct page *page = &folio->page;
 475         struct vm_area_struct *vma;
 476         unsigned long v_start;
 477         unsigned long v_end;
 478         pgoff_t start, end;
 479
 480         start = index * pages_per_huge_page(h);
 481         end = (index + 1) * pages_per_huge_page(h);
 482
 483         i_mmap_lock_write(mapping);
 484 retry:
 485         vma_lock = NULL;
 486         vma_interval_tree_foreach(vma, root, start, end - 1) {
 487                 v_start = vma_offset_start(vma, start);
 488                 v_end = vma_offset_end(vma, end);
 489
 490                 if (!hugetlb_vma_maps_page(vma, v_start, page))
 491                         continue;
 492
 493                 if (!hugetlb_vma_trylock_write(vma)) {
 494                         vma_lock = vma->vm_private_data;
 495                         /*
 496                          * If we can not get vma lock, we need to drop
 497                          * immap_sema and take locks in order.  First,
 498                          * take a ref on the vma_lock structure so that
 499                          * we can be guaranteed it will not go away when
 500                          * dropping immap_sema.
 501                          */
 502                         kref_get(&vma_lock->refs);
 503                         break;
 504                 }
 505
 506                 unmap_hugepage_range(vma, v_start, v_end, NULL,
 507                                      ZAP_FLAG_DROP_MARKER);
 508                 hugetlb_vma_unlock_write(vma);
 509         }
 510
 511         i_mmap_unlock_write(mapping);
 512
 513         if (vma_lock) {
 514                 /*
 515                  * Wait on vma_lock.  We know it is still valid as we have
 516                  * a reference.  We must 'open code' vma locking as we do
 517                  * not know if vma_lock is still attached to vma.
 518                  */
 519                 down_write(&vma_lock->rw_sema);
 520                 i_mmap_lock_write(mapping);
 521
 522                 vma = vma_lock->vma;
 523                 if (!vma) {
 524                         /*
 525                          * If lock is no longer attached to vma, then just
 526                          * unlock, drop our reference and retry looking for
 527                          * other vmas.
 528                          */
 529                         up_write(&vma_lock->rw_sema);
 530                         kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
 531                         goto retry;
 532                 }
 533
 534                 /*
 535                  * vma_lock is still attached to vma.  Check to see if vma
 536                  * still maps page and if so, unmap.
 537                  */
 538                 v_start = vma_offset_start(vma, start);
 539                 v_end = vma_offset_end(vma, end);
 540                 if (hugetlb_vma_maps_page(vma, v_start, page))
 541                         unmap_hugepage_range(vma, v_start, v_end, NULL,
 542                                              ZAP_FLAG_DROP_MARKER);
 543
 544                 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
 545                 hugetlb_vma_unlock_write(vma);
 546
 547                 goto retry;
 548         }
 549 }
 550
 551 static void
 552 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
 553                       zap_flags_t zap_flags)
 554 {
 555         struct vm_area_struct *vma;
 556
 557         /*
 558          * end == 0 indicates that the entire range after start should be
 559          * unmapped.  Note, end is exclusive, whereas the interval tree takes
 560          * an inclusive "last".
 561          */
 562         vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
 563                 unsigned long v_start;
 564                 unsigned long v_end;
 565
 566                 if (!hugetlb_vma_trylock_write(vma))
 567                         continue;
 568
 569                 v_start = vma_offset_start(vma, start);
 570                 v_end = vma_offset_end(vma, end);
 571
 572                 unmap_hugepage_range(vma, v_start, v_end, NULL, zap_flags);
 573
 574                 /*
 575                  * Note that vma lock only exists for shared/non-private
 576                  * vmas.  Therefore, lock is not held when calling
 577                  * unmap_hugepage_range for private vmas.
 578                  */
 579                 hugetlb_vma_unlock_write(vma);
 580         }
 581 }
 582
 583 /*
 584  * Called with hugetlb fault mutex held.
 585  * Returns true if page was actually removed, false otherwise.
 586  */
 587 static bool remove_inode_single_folio(struct hstate *h, struct inode *inode,
 588                                         struct address_space *mapping,
 589                                         struct folio *folio, pgoff_t index,
 590                                         bool truncate_op)
 591 {
 592         bool ret = false;
 593
 594         /*
 595          * If folio is mapped, it was faulted in after being
 596          * unmapped in caller.  Unmap (again) while holding
 597          * the fault mutex.  The mutex will prevent faults
 598          * until we finish removing the folio.
 599          */
 600         if (unlikely(folio_mapped(folio)))
 601                 hugetlb_unmap_file_folio(h, mapping, folio, index);
 602
 603         folio_lock(folio);
 604         /*
 605          * We must remove the folio from page cache before removing
 606          * the region/ reserve map (hugetlb_unreserve_pages).  In
 607          * rare out of memory conditions, removal of the region/reserve
 608          * map could fail.  Correspondingly, the subpool and global
 609          * reserve usage count can need to be adjusted.
 610          */
 611         VM_BUG_ON_FOLIO(folio_test_hugetlb_restore_reserve(folio), folio);
 612         hugetlb_delete_from_page_cache(folio);
 613         ret = true;
 614         if (!truncate_op) {
 615                 if (unlikely(hugetlb_unreserve_pages(inode, index,
 616                                                         index + 1, 1)))
 617                         hugetlb_fix_reserve_counts(inode);
 618         }
 619
 620         folio_unlock(folio);
 621         return ret;
 622 }
 623
 624 /*
 625  * remove_inode_hugepages handles two distinct cases: truncation and hole
 626  * punch.  There are subtle differences in operation for each case.
 627  *
 628  * truncation is indicated by end of range being LLONG_MAX
 629  *      In this case, we first scan the range and release found pages.
 630  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
 631  *      maps and global counts.  Page faults can race with truncation.
 632  *      During faults, hugetlb_no_page() checks i_size before page allocation,
 633  *      and again after obtaining page table lock.  It will 'back out'
 634  *      allocations in the truncated range.
 635  * hole punch is indicated if end is not LLONG_MAX
 636  *      In the hole punch case we scan the range and release found pages.
 637  *      Only when releasing a page is the associated region/reserve map
 638  *      deleted.  The region/reserve map for ranges without associated
 639  *      pages are not modified.  Page faults can race with hole punch.
 640  *      This is indicated if we find a mapped page.
 641  * Note: If the passed end of range value is beyond the end of file, but
 642  * not LLONG_MAX this routine still performs a hole punch operation.
 643  */
 644 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 645                                    loff_t lend)
 646 {
 647         struct hstate *h = hstate_inode(inode);
 648         struct address_space *mapping = &inode->i_data;
 649         const pgoff_t end = lend >> PAGE_SHIFT;
 650         struct folio_batch fbatch;
 651         pgoff_t next, index;
 652         int i, freed = 0;
 653         bool truncate_op = (lend == LLONG_MAX);
 654
 655         folio_batch_init(&fbatch);
 656         next = lstart >> PAGE_SHIFT;
 657         while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
 658                 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
 659                         struct folio *folio = fbatch.folios[i];
 660                         u32 hash = 0;
 661
 662                         index = folio->index >> huge_page_order(h);
 663                         hash = hugetlb_fault_mutex_hash(mapping, index);
 664                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 665
 666                         /*
 667                          * Remove folio that was part of folio_batch.
 668                          */
 669                         if (remove_inode_single_folio(h, inode, mapping, folio,
 670                                                         index, truncate_op))
 671                                 freed++;
 672
 673                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 674                 }
 675                 folio_batch_release(&fbatch);
 676                 cond_resched();
 677         }
 678
 679         if (truncate_op)
 680                 (void)hugetlb_unreserve_pages(inode,
 681                                 lstart >> huge_page_shift(h),
 682                                 LONG_MAX, freed);
 683 }
 684
 685 static void hugetlbfs_evict_inode(struct inode *inode)
 686 {
 687         struct resv_map *resv_map;
 688
 689         remove_inode_hugepages(inode, 0, LLONG_MAX);
 690
 691         /*
 692          * Get the resv_map from the address space embedded in the inode.
 693          * This is the address space which points to any resv_map allocated
 694          * at inode creation time.  If this is a device special inode,
 695          * i_mapping may not point to the original address space.
 696          */
 697         resv_map = (struct resv_map *)(&inode->i_data)->i_private_data;
 698         /* Only regular and link inodes have associated reserve maps */
 699         if (resv_map)
 700                 resv_map_release(&resv_map->refs);
 701         clear_inode(inode);
 702 }
 703
 704 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 705 {
 706         pgoff_t pgoff;
 707         struct address_space *mapping = inode->i_mapping;
 708         struct hstate *h = hstate_inode(inode);
 709
 710         BUG_ON(offset & ~huge_page_mask(h));
 711         pgoff = offset >> PAGE_SHIFT;
 712
 713         i_size_write(inode, offset);
 714         i_mmap_lock_write(mapping);
 715         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 716                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
 717                                       ZAP_FLAG_DROP_MARKER);
 718         i_mmap_unlock_write(mapping);
 719         remove_inode_hugepages(inode, offset, LLONG_MAX);
 720 }
 721
 722 static void hugetlbfs_zero_partial_page(struct hstate *h,
 723                                         struct address_space *mapping,
 724                                         loff_t start,
 725                                         loff_t end)
 726 {
 727         pgoff_t idx = start >> huge_page_shift(h);
 728         struct folio *folio;
 729
 730         folio = filemap_lock_hugetlb_folio(h, mapping, idx);
 731         if (IS_ERR(folio))
 732                 return;
 733
 734         start = start & ~huge_page_mask(h);
 735         end = end & ~huge_page_mask(h);
 736         if (!end)
 737                 end = huge_page_size(h);
 738
 739         folio_zero_segment(folio, (size_t)start, (size_t)end);
 740
 741         folio_unlock(folio);
 742         folio_put(folio);
 743 }
 744
 745 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 746 {
 747         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 748         struct address_space *mapping = inode->i_mapping;
 749         struct hstate *h = hstate_inode(inode);
 750         loff_t hpage_size = huge_page_size(h);
 751         loff_t hole_start, hole_end;
 752
 753         /*
 754          * hole_start and hole_end indicate the full pages within the hole.
 755          */
 756         hole_start = round_up(offset, hpage_size);
 757         hole_end = round_down(offset + len, hpage_size);
 758
 759         inode_lock(inode);
 760
 761         /* protected by i_rwsem */
 762         if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 763                 inode_unlock(inode);
 764                 return -EPERM;
 765         }
 766
 767         i_mmap_lock_write(mapping);
 768
 769         /* If range starts before first full page, zero partial page. */
 770         if (offset < hole_start)
 771                 hugetlbfs_zero_partial_page(h, mapping,
 772                                 offset, min(offset + len, hole_start));
 773
 774         /* Unmap users of full pages in the hole. */
 775         if (hole_end > hole_start) {
 776                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 777                         hugetlb_vmdelete_list(&mapping->i_mmap,
 778                                               hole_start >> PAGE_SHIFT,
 779                                               hole_end >> PAGE_SHIFT, 0);
 780         }
 781
 782         /* If range extends beyond last full page, zero partial page. */
 783         if ((offset + len) > hole_end && (offset + len) > hole_start)
 784                 hugetlbfs_zero_partial_page(h, mapping,
 785                                 hole_end, offset + len);
 786
 787         i_mmap_unlock_write(mapping);
 788
 789         /* Remove full pages from the file. */
 790         if (hole_end > hole_start)
 791                 remove_inode_hugepages(inode, hole_start, hole_end);
 792
 793         inode_unlock(inode);
 794
 795         return 0;
 796 }
 797
 798 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 799                                 loff_t len)
 800 {
 801         struct inode *inode = file_inode(file);
 802         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 803         struct address_space *mapping = inode->i_mapping;
 804         struct hstate *h = hstate_inode(inode);
 805         struct vm_area_struct pseudo_vma;
 806         struct mm_struct *mm = current->mm;
 807         loff_t hpage_size = huge_page_size(h);
 808         unsigned long hpage_shift = huge_page_shift(h);
 809         pgoff_t start, index, end;
 810         int error;
 811         u32 hash;
 812
 813         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 814                 return -EOPNOTSUPP;
 815
 816         if (mode & FALLOC_FL_PUNCH_HOLE)
 817                 return hugetlbfs_punch_hole(inode, offset, len);
 818
 819         /*
 820          * Default preallocate case.
 821          * For this range, start is rounded down and end is rounded up
 822          * as well as being converted to page offsets.
 823          */
 824         start = offset >> hpage_shift;
 825         end = (offset + len + hpage_size - 1) >> hpage_shift;
 826
 827         inode_lock(inode);
 828
 829         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 830         error = inode_newsize_ok(inode, offset + len);
 831         if (error)
 832                 goto out;
 833
 834         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 835                 error = -EPERM;
 836                 goto out;
 837         }
 838
 839         /*
 840          * Initialize a pseudo vma as this is required by the huge page
 841          * allocation routines.
 842          */
 843         vma_init(&pseudo_vma, mm);
 844         vm_flags_init(&pseudo_vma, VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 845         pseudo_vma.vm_file = file;
 846
 847         for (index = start; index < end; index++) {
 848                 /*
 849                  * This is supposed to be the vaddr where the page is being
 850                  * faulted in, but we have no vaddr here.
 851                  */
 852                 struct folio *folio;
 853                 unsigned long addr;
 854
 855                 cond_resched();
 856
 857                 /*
 858                  * fallocate(2) manpage permits EINTR; we may have been
 859                  * interrupted because we are using up too much memory.
 860                  */
 861                 if (signal_pending(current)) {
 862                         error = -EINTR;
 863                         break;
 864                 }
 865
 866                 /* addr is the offset within the file (zero based) */
 867                 addr = index * hpage_size;
 868
 869                 /* mutex taken here, fault path and hole punch */
 870                 hash = hugetlb_fault_mutex_hash(mapping, index);
 871                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 872
 873                 /* See if already present in mapping to avoid alloc/free */
 874                 folio = filemap_get_folio(mapping, index << huge_page_order(h));
 875                 if (!IS_ERR(folio)) {
 876                         folio_put(folio);
 877                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 878                         continue;
 879                 }
 880
 881                 /*
 882                  * Allocate folio without setting the avoid_reserve argument.
 883                  * There certainly are no reserves associated with the
 884                  * pseudo_vma.  However, there could be shared mappings with
 885                  * reserves for the file at the inode level.  If we fallocate
 886                  * folios in these areas, we need to consume the reserves
 887                  * to keep reservation accounting consistent.
 888                  */
 889                 folio = alloc_hugetlb_folio(&pseudo_vma, addr, 0);
 890                 if (IS_ERR(folio)) {
 891                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 892                         error = PTR_ERR(folio);
 893                         goto out;
 894                 }
 895                 clear_huge_page(&folio->page, addr, pages_per_huge_page(h));
 896                 __folio_mark_uptodate(folio);
 897                 error = hugetlb_add_to_page_cache(folio, mapping, index);
 898                 if (unlikely(error)) {
 899                         restore_reserve_on_error(h, &pseudo_vma, addr, folio);
 900                         folio_put(folio);
 901                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 902                         goto out;
 903                 }
 904
 905                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 906
 907                 folio_set_hugetlb_migratable(folio);
 908                 /*
 909                  * folio_unlock because locked by hugetlb_add_to_page_cache()
 910                  * folio_put() due to reference from alloc_hugetlb_folio()
 911                  */
 912                 folio_unlock(folio);
 913                 folio_put(folio);
 914         }
 915
 916         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 917                 i_size_write(inode, offset + len);
 918         inode_set_ctime_current(inode);
 919 out:
 920         inode_unlock(inode);
 921         return error;
 922 }
 923
 924 static int hugetlbfs_setattr(struct mnt_idmap *idmap,
 925                              struct dentry *dentry, struct iattr *attr)
 926 {
 927         struct inode *inode = d_inode(dentry);
 928         struct hstate *h = hstate_inode(inode);
 929         int error;
 930         unsigned int ia_valid = attr->ia_valid;
 931         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 932
 933         error = setattr_prepare(idmap, dentry, attr);
 934         if (error)
 935                 return error;
 936
 937         if (ia_valid & ATTR_SIZE) {
 938                 loff_t oldsize = inode->i_size;
 939                 loff_t newsize = attr->ia_size;
 940
 941                 if (newsize & ~huge_page_mask(h))
 942                         return -EINVAL;
 943                 /* protected by i_rwsem */
 944                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 945                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 946                         return -EPERM;
 947                 hugetlb_vmtruncate(inode, newsize);
 948         }
 949
 950         setattr_copy(idmap, inode, attr);
 951         mark_inode_dirty(inode);
 952         return 0;
 953 }
 954
 955 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 956                                         struct hugetlbfs_fs_context *ctx)
 957 {
 958         struct inode *inode;
 959
 960         inode = new_inode(sb);
 961         if (inode) {
 962                 inode->i_ino = get_next_ino();
 963                 inode->i_mode = S_IFDIR | ctx->mode;
 964                 inode->i_uid = ctx->uid;
 965                 inode->i_gid = ctx->gid;
 966                 simple_inode_init_ts(inode);
 967                 inode->i_op = &hugetlbfs_dir_inode_operations;
 968                 inode->i_fop = &simple_dir_operations;
 969                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 970                 inc_nlink(inode);
 971                 lockdep_annotate_inode_mutex_key(inode);
 972         }
 973         return inode;
 974 }
 975
 976 /*
 977  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 978  * be taken from reclaim -- unlike regular filesystems. This needs an
 979  * annotation because huge_pmd_share() does an allocation under hugetlb's
 980  * i_mmap_rwsem.
 981  */
 982 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 983
 984 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 985                                         struct mnt_idmap *idmap,
 986                                         struct inode *dir,
 987                                         umode_t mode, dev_t dev)
 988 {
 989         struct inode *inode;
 990         struct resv_map *resv_map = NULL;
 991
 992         /*
 993          * Reserve maps are only needed for inodes that can have associated
 994          * page allocations.
 995          */
 996         if (S_ISREG(mode) || S_ISLNK(mode)) {
 997                 resv_map = resv_map_alloc();
 998                 if (!resv_map)
 999                         return NULL;
1000         }
1001
1002         inode = new_inode(sb);
1003         if (inode) {
1004                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
1005
1006                 inode->i_ino = get_next_ino();
1007                 inode_init_owner(idmap, inode, dir, mode);
1008                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
1009                                 &hugetlbfs_i_mmap_rwsem_key);
1010                 inode->i_mapping->a_ops = &hugetlbfs_aops;
1011                 simple_inode_init_ts(inode);
1012                 inode->i_mapping->i_private_data = resv_map;
1013                 info->seals = F_SEAL_SEAL;
1014                 switch (mode & S_IFMT) {
1015                 default:
1016                         init_special_inode(inode, mode, dev);
1017                         break;
1018                 case S_IFREG:
1019                         inode->i_op = &hugetlbfs_inode_operations;
1020                         inode->i_fop = &hugetlbfs_file_operations;
1021                         break;
1022                 case S_IFDIR:
1023                         inode->i_op = &hugetlbfs_dir_inode_operations;
1024                         inode->i_fop = &simple_dir_operations;
1025
1026                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
1027                         inc_nlink(inode);
1028                         break;
1029                 case S_IFLNK:
1030                         inode->i_op = &page_symlink_inode_operations;
1031                         inode_nohighmem(inode);
1032                         break;
1033                 }
1034                 lockdep_annotate_inode_mutex_key(inode);
1035         } else {
1036                 if (resv_map)
1037                         kref_put(&resv_map->refs, resv_map_release);
1038         }
1039
1040         return inode;
1041 }
1042
1043 /*
1044  * File creation. Allocate an inode, and we're done..
1045  */
1046 static int hugetlbfs_mknod(struct mnt_idmap *idmap, struct inode *dir,
1047                            struct dentry *dentry, umode_t mode, dev_t dev)
1048 {
1049         struct inode *inode;
1050
1051         inode = hugetlbfs_get_inode(dir->i_sb, idmap, dir, mode, dev);
1052         if (!inode)
1053                 return -ENOSPC;
1054         inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1055         d_instantiate(dentry, inode);
1056         dget(dentry);/* Extra count - pin the dentry in core */
1057         return 0;
1058 }
1059
1060 static int hugetlbfs_mkdir(struct mnt_idmap *idmap, struct inode *dir,
1061                            struct dentry *dentry, umode_t mode)
1062 {
1063         int retval = hugetlbfs_mknod(idmap, dir, dentry,
1064                                      mode | S_IFDIR, 0);
1065         if (!retval)
1066                 inc_nlink(dir);
1067         return retval;
1068 }
1069
1070 static int hugetlbfs_create(struct mnt_idmap *idmap,
1071                             struct inode *dir, struct dentry *dentry,
1072                             umode_t mode, bool excl)
1073 {
1074         return hugetlbfs_mknod(idmap, dir, dentry, mode | S_IFREG, 0);
1075 }
1076
1077 static int hugetlbfs_tmpfile(struct mnt_idmap *idmap,
1078                              struct inode *dir, struct file *file,
1079                              umode_t mode)
1080 {
1081         struct inode *inode;
1082
1083         inode = hugetlbfs_get_inode(dir->i_sb, idmap, dir, mode | S_IFREG, 0);
1084         if (!inode)
1085                 return -ENOSPC;
1086         inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1087         d_tmpfile(file, inode);
1088         return finish_open_simple(file, 0);
1089 }
1090
1091 static int hugetlbfs_symlink(struct mnt_idmap *idmap,
1092                              struct inode *dir, struct dentry *dentry,
1093                              const char *symname)
1094 {
1095         const umode_t mode = S_IFLNK|S_IRWXUGO;
1096         struct inode *inode;
1097         int error = -ENOSPC;
1098
1099         inode = hugetlbfs_get_inode(dir->i_sb, idmap, dir, mode, 0);
1100         if (inode) {
1101                 int l = strlen(symname)+1;
1102                 error = page_symlink(inode, symname, l);
1103                 if (!error) {
1104                         d_instantiate(dentry, inode);
1105                         dget(dentry);
1106                 } else
1107                         iput(inode);
1108         }
1109         inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1110
1111         return error;
1112 }
1113
1114 #ifdef CONFIG_MIGRATION
1115 static int hugetlbfs_migrate_folio(struct address_space *mapping,
1116                                 struct folio *dst, struct folio *src,
1117                                 enum migrate_mode mode)
1118 {
1119         int rc;
1120
1121         rc = migrate_huge_page_move_mapping(mapping, dst, src);
1122         if (rc != MIGRATEPAGE_SUCCESS)
1123                 return rc;
1124
1125         if (hugetlb_folio_subpool(src)) {
1126                 hugetlb_set_folio_subpool(dst,
1127                                         hugetlb_folio_subpool(src));
1128                 hugetlb_set_folio_subpool(src, NULL);
1129         }
1130
1131         if (mode != MIGRATE_SYNC_NO_COPY)
1132                 folio_migrate_copy(dst, src);
1133         else
1134                 folio_migrate_flags(dst, src);
1135
1136         return MIGRATEPAGE_SUCCESS;
1137 }
1138 #else
1139 #define hugetlbfs_migrate_folio NULL
1140 #endif
1141
1142 static int hugetlbfs_error_remove_folio(struct address_space *mapping,
1143                                 struct folio *folio)
1144 {
1145         return 0;
1146 }
1147
1148 /*
1149  * Display the mount options in /proc/mounts.
1150  */
1151 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1152 {
1153         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1154         struct hugepage_subpool *spool = sbinfo->spool;
1155         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1156         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1157         char mod;
1158
1159         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1160                 seq_printf(m, ",uid=%u",
1161                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1162         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1163                 seq_printf(m, ",gid=%u",
1164                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1165         if (sbinfo->mode != 0755)
1166                 seq_printf(m, ",mode=%o", sbinfo->mode);
1167         if (sbinfo->max_inodes != -1)
1168                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1169
1170         hpage_size /= 1024;
1171         mod = 'K';
1172         if (hpage_size >= 1024) {
1173                 hpage_size /= 1024;
1174                 mod = 'M';
1175         }
1176         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1177         if (spool) {
1178                 if (spool->max_hpages != -1)
1179                         seq_printf(m, ",size=%llu",
1180                                    (unsigned long long)spool->max_hpages << hpage_shift);
1181                 if (spool->min_hpages != -1)
1182                         seq_printf(m, ",min_size=%llu",
1183                                    (unsigned long long)spool->min_hpages << hpage_shift);
1184         }
1185         return 0;
1186 }
1187
1188 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1189 {
1190         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1191         struct hstate *h = hstate_inode(d_inode(dentry));
1192         u64 id = huge_encode_dev(dentry->d_sb->s_dev);
1193
1194         buf->f_fsid = u64_to_fsid(id);
1195         buf->f_type = HUGETLBFS_MAGIC;
1196         buf->f_bsize = huge_page_size(h);
1197         if (sbinfo) {
1198                 spin_lock(&sbinfo->stat_lock);
1199                 /* If no limits set, just report 0 or -1 for max/free/used
1200                  * blocks, like simple_statfs() */
1201                 if (sbinfo->spool) {
1202                         long free_pages;
1203
1204                         spin_lock_irq(&sbinfo->spool->lock);
1205                         buf->f_blocks = sbinfo->spool->max_hpages;
1206                         free_pages = sbinfo->spool->max_hpages
1207                                 - sbinfo->spool->used_hpages;
1208                         buf->f_bavail = buf->f_bfree = free_pages;
1209                         spin_unlock_irq(&sbinfo->spool->lock);
1210                         buf->f_files = sbinfo->max_inodes;
1211                         buf->f_ffree = sbinfo->free_inodes;
1212                 }
1213                 spin_unlock(&sbinfo->stat_lock);
1214         }
1215         buf->f_namelen = NAME_MAX;
1216         return 0;
1217 }
1218
1219 static void hugetlbfs_put_super(struct super_block *sb)
1220 {
1221         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1222
1223         if (sbi) {
1224                 sb->s_fs_info = NULL;
1225
1226                 if (sbi->spool)
1227                         hugepage_put_subpool(sbi->spool);
1228
1229                 kfree(sbi);
1230         }
1231 }
1232
1233 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1234 {
1235         if (sbinfo->free_inodes >= 0) {
1236                 spin_lock(&sbinfo->stat_lock);
1237                 if (unlikely(!sbinfo->free_inodes)) {
1238                         spin_unlock(&sbinfo->stat_lock);
1239                         return 0;
1240                 }
1241                 sbinfo->free_inodes--;
1242                 spin_unlock(&sbinfo->stat_lock);
1243         }
1244
1245         return 1;
1246 }
1247
1248 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1249 {
1250         if (sbinfo->free_inodes >= 0) {
1251                 spin_lock(&sbinfo->stat_lock);
1252                 sbinfo->free_inodes++;
1253                 spin_unlock(&sbinfo->stat_lock);
1254         }
1255 }
1256
1257
1258 static struct kmem_cache *hugetlbfs_inode_cachep;
1259
1260 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1261 {
1262         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1263         struct hugetlbfs_inode_info *p;
1264
1265         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1266                 return NULL;
1267         p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1268         if (unlikely(!p)) {
1269                 hugetlbfs_inc_free_inodes(sbinfo);
1270                 return NULL;
1271         }
1272         return &p->vfs_inode;
1273 }
1274
1275 static void hugetlbfs_free_inode(struct inode *inode)
1276 {
1277         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1278 }
1279
1280 static void hugetlbfs_destroy_inode(struct inode *inode)
1281 {
1282         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1283 }
1284
1285 static const struct address_space_operations hugetlbfs_aops = {
1286         .write_begin    = hugetlbfs_write_begin,
1287         .write_end      = hugetlbfs_write_end,
1288         .dirty_folio    = noop_dirty_folio,
1289         .migrate_folio  = hugetlbfs_migrate_folio,
1290         .error_remove_folio     = hugetlbfs_error_remove_folio,
1291 };
1292
1293
1294 static void init_once(void *foo)
1295 {
1296         struct hugetlbfs_inode_info *ei = foo;
1297
1298         inode_init_once(&ei->vfs_inode);
1299 }
1300
1301 static const struct file_operations hugetlbfs_file_operations = {
1302         .read_iter              = hugetlbfs_read_iter,
1303         .mmap                   = hugetlbfs_file_mmap,
1304         .fsync                  = noop_fsync,
1305         .get_unmapped_area      = hugetlb_get_unmapped_area,
1306         .llseek                 = default_llseek,
1307         .fallocate              = hugetlbfs_fallocate,
1308         .fop_flags              = FOP_HUGE_PAGES,
1309 };
1310
1311 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1312         .create         = hugetlbfs_create,
1313         .lookup         = simple_lookup,
1314         .link           = simple_link,
1315         .unlink         = simple_unlink,
1316         .symlink        = hugetlbfs_symlink,
1317         .mkdir          = hugetlbfs_mkdir,
1318         .rmdir          = simple_rmdir,
1319         .mknod          = hugetlbfs_mknod,
1320         .rename         = simple_rename,
1321         .setattr        = hugetlbfs_setattr,
1322         .tmpfile        = hugetlbfs_tmpfile,
1323 };
1324
1325 static const struct inode_operations hugetlbfs_inode_operations = {
1326         .setattr        = hugetlbfs_setattr,
1327 };
1328
1329 static const struct super_operations hugetlbfs_ops = {
1330         .alloc_inode    = hugetlbfs_alloc_inode,
1331         .free_inode     = hugetlbfs_free_inode,
1332         .destroy_inode  = hugetlbfs_destroy_inode,
1333         .evict_inode    = hugetlbfs_evict_inode,
1334         .statfs         = hugetlbfs_statfs,
1335         .put_super      = hugetlbfs_put_super,
1336         .show_options   = hugetlbfs_show_options,
1337 };
1338
1339 /*
1340  * Convert size option passed from command line to number of huge pages
1341  * in the pool specified by hstate.  Size option could be in bytes
1342  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1343  */
1344 static long
1345 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1346                          enum hugetlbfs_size_type val_type)
1347 {
1348         if (val_type == NO_SIZE)
1349                 return -1;
1350
1351         if (val_type == SIZE_PERCENT) {
1352                 size_opt <<= huge_page_shift(h);
1353                 size_opt *= h->max_huge_pages;
1354                 do_div(size_opt, 100);
1355         }
1356
1357         size_opt >>= huge_page_shift(h);
1358         return size_opt;
1359 }
1360
1361 /*
1362  * Parse one mount parameter.
1363  */
1364 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1365 {
1366         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1367         struct fs_parse_result result;
1368         struct hstate *h;
1369         char *rest;
1370         unsigned long ps;
1371         int opt;
1372
1373         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1374         if (opt < 0)
1375                 return opt;
1376
1377         switch (opt) {
1378         case Opt_uid:
1379                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1380                 if (!uid_valid(ctx->uid))
1381                         goto bad_val;
1382                 return 0;
1383
1384         case Opt_gid:
1385                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1386                 if (!gid_valid(ctx->gid))
1387                         goto bad_val;
1388                 return 0;
1389
1390         case Opt_mode:
1391                 ctx->mode = result.uint_32 & 01777U;
1392                 return 0;
1393
1394         case Opt_size:
1395                 /* memparse() will accept a K/M/G without a digit */
1396                 if (!param->string || !isdigit(param->string[0]))
1397                         goto bad_val;
1398                 ctx->max_size_opt = memparse(param->string, &rest);
1399                 ctx->max_val_type = SIZE_STD;
1400                 if (*rest == '%')
1401                         ctx->max_val_type = SIZE_PERCENT;
1402                 return 0;
1403
1404         case Opt_nr_inodes:
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->nr_inodes = memparse(param->string, &rest);
1409                 return 0;
1410
1411         case Opt_pagesize:
1412                 ps = memparse(param->string, &rest);
1413                 h = size_to_hstate(ps);
1414                 if (!h) {
1415                         pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1416                         return -EINVAL;
1417                 }
1418                 ctx->hstate = h;
1419                 return 0;
1420
1421         case Opt_min_size:
1422                 /* memparse() will accept a K/M/G without a digit */
1423                 if (!param->string || !isdigit(param->string[0]))
1424                         goto bad_val;
1425                 ctx->min_size_opt = memparse(param->string, &rest);
1426                 ctx->min_val_type = SIZE_STD;
1427                 if (*rest == '%')
1428                         ctx->min_val_type = SIZE_PERCENT;
1429                 return 0;
1430
1431         default:
1432                 return -EINVAL;
1433         }
1434
1435 bad_val:
1436         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1437                       param->string, param->key);
1438 }
1439
1440 /*
1441  * Validate the parsed options.
1442  */
1443 static int hugetlbfs_validate(struct fs_context *fc)
1444 {
1445         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1446
1447         /*
1448          * Use huge page pool size (in hstate) to convert the size
1449          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1450          */
1451         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1452                                                    ctx->max_size_opt,
1453                                                    ctx->max_val_type);
1454         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1455                                                    ctx->min_size_opt,
1456                                                    ctx->min_val_type);
1457
1458         /*
1459          * If max_size was specified, then min_size must be smaller
1460          */
1461         if (ctx->max_val_type > NO_SIZE &&
1462             ctx->min_hpages > ctx->max_hpages) {
1463                 pr_err("Minimum size can not be greater than maximum size\n");
1464                 return -EINVAL;
1465         }
1466
1467         return 0;
1468 }
1469
1470 static int
1471 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1472 {
1473         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1474         struct hugetlbfs_sb_info *sbinfo;
1475
1476         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1477         if (!sbinfo)
1478                 return -ENOMEM;
1479         sb->s_fs_info = sbinfo;
1480         spin_lock_init(&sbinfo->stat_lock);
1481         sbinfo->hstate          = ctx->hstate;
1482         sbinfo->max_inodes      = ctx->nr_inodes;
1483         sbinfo->free_inodes     = ctx->nr_inodes;
1484         sbinfo->spool           = NULL;
1485         sbinfo->uid             = ctx->uid;
1486         sbinfo->gid             = ctx->gid;
1487         sbinfo->mode            = ctx->mode;
1488
1489         /*
1490          * Allocate and initialize subpool if maximum or minimum size is
1491          * specified.  Any needed reservations (for minimum size) are taken
1492          * when the subpool is created.
1493          */
1494         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1495                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1496                                                      ctx->max_hpages,
1497                                                      ctx->min_hpages);
1498                 if (!sbinfo->spool)
1499                         goto out_free;
1500         }
1501         sb->s_maxbytes = MAX_LFS_FILESIZE;
1502         sb->s_blocksize = huge_page_size(ctx->hstate);
1503         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1504         sb->s_magic = HUGETLBFS_MAGIC;
1505         sb->s_op = &hugetlbfs_ops;
1506         sb->s_time_gran = 1;
1507
1508         /*
1509          * Due to the special and limited functionality of hugetlbfs, it does
1510          * not work well as a stacking filesystem.
1511          */
1512         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1513         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1514         if (!sb->s_root)
1515                 goto out_free;
1516         return 0;
1517 out_free:
1518         kfree(sbinfo->spool);
1519         kfree(sbinfo);
1520         return -ENOMEM;
1521 }
1522
1523 static int hugetlbfs_get_tree(struct fs_context *fc)
1524 {
1525         int err = hugetlbfs_validate(fc);
1526         if (err)
1527                 return err;
1528         return get_tree_nodev(fc, hugetlbfs_fill_super);
1529 }
1530
1531 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1532 {
1533         kfree(fc->fs_private);
1534 }
1535
1536 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1537         .free           = hugetlbfs_fs_context_free,
1538         .parse_param    = hugetlbfs_parse_param,
1539         .get_tree       = hugetlbfs_get_tree,
1540 };
1541
1542 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1543 {
1544         struct hugetlbfs_fs_context *ctx;
1545
1546         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1547         if (!ctx)
1548                 return -ENOMEM;
1549
1550         ctx->max_hpages = -1; /* No limit on size by default */
1551         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1552         ctx->uid        = current_fsuid();
1553         ctx->gid        = current_fsgid();
1554         ctx->mode       = 0755;
1555         ctx->hstate     = &default_hstate;
1556         ctx->min_hpages = -1; /* No default minimum size */
1557         ctx->max_val_type = NO_SIZE;
1558         ctx->min_val_type = NO_SIZE;
1559         fc->fs_private = ctx;
1560         fc->ops = &hugetlbfs_fs_context_ops;
1561         return 0;
1562 }
1563
1564 static struct file_system_type hugetlbfs_fs_type = {
1565         .name                   = "hugetlbfs",
1566         .init_fs_context        = hugetlbfs_init_fs_context,
1567         .parameters             = hugetlb_fs_parameters,
1568         .kill_sb                = kill_litter_super,
1569         .fs_flags               = FS_ALLOW_IDMAP,
1570 };
1571
1572 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1573
1574 static int can_do_hugetlb_shm(void)
1575 {
1576         kgid_t shm_group;
1577         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1578         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1579 }
1580
1581 static int get_hstate_idx(int page_size_log)
1582 {
1583         struct hstate *h = hstate_sizelog(page_size_log);
1584
1585         if (!h)
1586                 return -1;
1587         return hstate_index(h);
1588 }
1589
1590 /*
1591  * Note that size should be aligned to proper hugepage size in caller side,
1592  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1593  */
1594 struct file *hugetlb_file_setup(const char *name, size_t size,
1595                                 vm_flags_t acctflag, int creat_flags,
1596                                 int page_size_log)
1597 {
1598         struct inode *inode;
1599         struct vfsmount *mnt;
1600         int hstate_idx;
1601         struct file *file;
1602
1603         hstate_idx = get_hstate_idx(page_size_log);
1604         if (hstate_idx < 0)
1605                 return ERR_PTR(-ENODEV);
1606
1607         mnt = hugetlbfs_vfsmount[hstate_idx];
1608         if (!mnt)
1609                 return ERR_PTR(-ENOENT);
1610
1611         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1612                 struct ucounts *ucounts = current_ucounts();
1613
1614                 if (user_shm_lock(size, ucounts)) {
1615                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1616                                 current->comm, current->pid);
1617                         user_shm_unlock(size, ucounts);
1618                 }
1619                 return ERR_PTR(-EPERM);
1620         }
1621
1622         file = ERR_PTR(-ENOSPC);
1623         /* hugetlbfs_vfsmount[] mounts do not use idmapped mounts.  */
1624         inode = hugetlbfs_get_inode(mnt->mnt_sb, &nop_mnt_idmap, NULL,
1625                                     S_IFREG | S_IRWXUGO, 0);
1626         if (!inode)
1627                 goto out;
1628         if (creat_flags == HUGETLB_SHMFS_INODE)
1629                 inode->i_flags |= S_PRIVATE;
1630
1631         inode->i_size = size;
1632         clear_nlink(inode);
1633
1634         if (!hugetlb_reserve_pages(inode, 0,
1635                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1636                         acctflag))
1637                 file = ERR_PTR(-ENOMEM);
1638         else
1639                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1640                                         &hugetlbfs_file_operations);
1641         if (!IS_ERR(file))
1642                 return file;
1643
1644         iput(inode);
1645 out:
1646         return file;
1647 }
1648
1649 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1650 {
1651         struct fs_context *fc;
1652         struct vfsmount *mnt;
1653
1654         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1655         if (IS_ERR(fc)) {
1656                 mnt = ERR_CAST(fc);
1657         } else {
1658                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1659                 ctx->hstate = h;
1660                 mnt = fc_mount(fc);
1661                 put_fs_context(fc);
1662         }
1663         if (IS_ERR(mnt))
1664                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1665                        huge_page_size(h) / SZ_1K);
1666         return mnt;
1667 }
1668
1669 static int __init init_hugetlbfs_fs(void)
1670 {
1671         struct vfsmount *mnt;
1672         struct hstate *h;
1673         int error;
1674         int i;
1675
1676         if (!hugepages_supported()) {
1677                 pr_info("disabling because there are no supported hugepage sizes\n");
1678                 return -ENOTSUPP;
1679         }
1680
1681         error = -ENOMEM;
1682         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1683                                         sizeof(struct hugetlbfs_inode_info),
1684                                         0, SLAB_ACCOUNT, init_once);
1685         if (hugetlbfs_inode_cachep == NULL)
1686                 goto out;
1687
1688         error = register_filesystem(&hugetlbfs_fs_type);
1689         if (error)
1690                 goto out_free;
1691
1692         /* default hstate mount is required */
1693         mnt = mount_one_hugetlbfs(&default_hstate);
1694         if (IS_ERR(mnt)) {
1695                 error = PTR_ERR(mnt);
1696                 goto out_unreg;
1697         }
1698         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1699
1700         /* other hstates are optional */
1701         i = 0;
1702         for_each_hstate(h) {
1703                 if (i == default_hstate_idx) {
1704                         i++;
1705                         continue;
1706                 }
1707
1708                 mnt = mount_one_hugetlbfs(h);
1709                 if (IS_ERR(mnt))
1710                         hugetlbfs_vfsmount[i] = NULL;
1711                 else
1712                         hugetlbfs_vfsmount[i] = mnt;
1713                 i++;
1714         }
1715
1716         return 0;
1717
1718  out_unreg:
1719         (void)unregister_filesystem(&hugetlbfs_fs_type);
1720  out_free:
1721         kmem_cache_destroy(hugetlbfs_inode_cachep);
1722  out:
1723         return error;
1724 }
1725 fs_initcall(init_hugetlbfs_fs)