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 /*
  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.flags = 0;
 182         info.length = len;
 183         info.low_limit = current->mm->mmap_base;
 184         info.high_limit = arch_get_mmap_end(addr, len, flags);
 185         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 186         info.align_offset = 0;
 187         return vm_unmapped_area(&info);
 188 }
 189
 190 static unsigned long
 191 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
 192                 unsigned long len, unsigned long pgoff, unsigned long flags)
 193 {
 194         struct hstate *h = hstate_file(file);
 195         struct vm_unmapped_area_info info;
 196
 197         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 198         info.length = len;
 199         info.low_limit = PAGE_SIZE;
 200         info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
 201         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 202         info.align_offset = 0;
 203         addr = vm_unmapped_area(&info);
 204
 205         /*
 206          * A failed mmap() very likely causes application failure,
 207          * so fall back to the bottom-up function here. This scenario
 208          * can happen with large stack limits and large mmap()
 209          * allocations.
 210          */
 211         if (unlikely(offset_in_page(addr))) {
 212                 VM_BUG_ON(addr != -ENOMEM);
 213                 info.flags = 0;
 214                 info.low_limit = current->mm->mmap_base;
 215                 info.high_limit = arch_get_mmap_end(addr, len, flags);
 216                 addr = vm_unmapped_area(&info);
 217         }
 218
 219         return addr;
 220 }
 221
 222 unsigned long
 223 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 224                                   unsigned long len, unsigned long pgoff,
 225                                   unsigned long flags)
 226 {
 227         struct mm_struct *mm = current->mm;
 228         struct vm_area_struct *vma;
 229         struct hstate *h = hstate_file(file);
 230         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
 231
 232         if (len & ~huge_page_mask(h))
 233                 return -EINVAL;
 234         if (len > TASK_SIZE)
 235                 return -ENOMEM;
 236
 237         if (flags & MAP_FIXED) {
 238                 if (prepare_hugepage_range(file, addr, len))
 239                         return -EINVAL;
 240                 return addr;
 241         }
 242
 243         if (addr) {
 244                 addr = ALIGN(addr, huge_page_size(h));
 245                 vma = find_vma(mm, addr);
 246                 if (mmap_end - len >= addr &&
 247                     (!vma || addr + len <= vm_start_gap(vma)))
 248                         return addr;
 249         }
 250
 251         /*
 252          * Use mm->get_unmapped_area value as a hint to use topdown routine.
 253          * If architectures have special needs, they should define their own
 254          * version of hugetlb_get_unmapped_area.
 255          */
 256         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
 257                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 258                                 pgoff, flags);
 259         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 260                         pgoff, flags);
 261 }
 262
 263 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 264 static unsigned long
 265 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 266                           unsigned long len, unsigned long pgoff,
 267                           unsigned long flags)
 268 {
 269         return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
 270 }
 271 #endif
 272
 273 /*
 274  * Someone wants to read @bytes from a HWPOISON hugetlb @page from @offset.
 275  * Returns the maximum number of bytes one can read without touching the 1st raw
 276  * HWPOISON subpage.
 277  *
 278  * The implementation borrows the iteration logic from copy_page_to_iter*.
 279  */
 280 static size_t adjust_range_hwpoison(struct page *page, size_t offset, size_t bytes)
 281 {
 282         size_t n = 0;
 283         size_t res = 0;
 284
 285         /* First subpage to start the loop. */
 286         page = nth_page(page, offset / PAGE_SIZE);
 287         offset %= PAGE_SIZE;
 288         while (1) {
 289                 if (is_raw_hwpoison_page_in_hugepage(page))
 290                         break;
 291
 292                 /* Safe to read n bytes without touching HWPOISON subpage. */
 293                 n = min(bytes, (size_t)PAGE_SIZE - offset);
 294                 res += n;
 295                 bytes -= n;
 296                 if (!bytes || !n)
 297                         break;
 298                 offset += n;
 299                 if (offset == PAGE_SIZE) {
 300                         page = nth_page(page, 1);
 301                         offset = 0;
 302                 }
 303         }
 304
 305         return res;
 306 }
 307
 308 /*
 309  * Support for read() - Find the page attached to f_mapping and copy out the
 310  * data. This provides functionality similar to filemap_read().
 311  */
 312 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 313 {
 314         struct file *file = iocb->ki_filp;
 315         struct hstate *h = hstate_file(file);
 316         struct address_space *mapping = file->f_mapping;
 317         struct inode *inode = mapping->host;
 318         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 319         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 320         unsigned long end_index;
 321         loff_t isize;
 322         ssize_t retval = 0;
 323
 324         while (iov_iter_count(to)) {
 325                 struct folio *folio;
 326                 size_t nr, copied, want;
 327
 328                 /* nr is the maximum number of bytes to copy from this page */
 329                 nr = huge_page_size(h);
 330                 isize = i_size_read(inode);
 331                 if (!isize)
 332                         break;
 333                 end_index = (isize - 1) >> huge_page_shift(h);
 334                 if (index > end_index)
 335                         break;
 336                 if (index == end_index) {
 337                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 338                         if (nr <= offset)
 339                                 break;
 340                 }
 341                 nr = nr - offset;
 342
 343                 /* Find the folio */
 344                 folio = filemap_lock_hugetlb_folio(h, mapping, index);
 345                 if (IS_ERR(folio)) {
 346                         /*
 347                          * We have a HOLE, zero out the user-buffer for the
 348                          * length of the hole or request.
 349                          */
 350                         copied = iov_iter_zero(nr, to);
 351                 } else {
 352                         folio_unlock(folio);
 353
 354                         if (!folio_test_hwpoison(folio))
 355                                 want = nr;
 356                         else {
 357                                 /*
 358                                  * Adjust how many bytes safe to read without
 359                                  * touching the 1st raw HWPOISON subpage after
 360                                  * offset.
 361                                  */
 362                                 want = adjust_range_hwpoison(&folio->page, offset, nr);
 363                                 if (want == 0) {
 364                                         folio_put(folio);
 365                                         retval = -EIO;
 366                                         break;
 367                                 }
 368                         }
 369
 370                         /*
 371                          * We have the folio, copy it to user space buffer.
 372                          */
 373                         copied = copy_folio_to_iter(folio, offset, want, to);
 374                         folio_put(folio);
 375                 }
 376                 offset += copied;
 377                 retval += copied;
 378                 if (copied != nr && iov_iter_count(to)) {
 379                         if (!retval)
 380                                 retval = -EFAULT;
 381                         break;
 382                 }
 383                 index += offset >> huge_page_shift(h);
 384                 offset &= ~huge_page_mask(h);
 385         }
 386         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 387         return retval;
 388 }
 389
 390 static int hugetlbfs_write_begin(struct file *file,
 391                         struct address_space *mapping,
 392                         loff_t pos, unsigned len,
 393                         struct page **pagep, void **fsdata)
 394 {
 395         return -EINVAL;
 396 }
 397
 398 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 399                         loff_t pos, unsigned len, unsigned copied,
 400                         struct page *page, void *fsdata)
 401 {
 402         BUG();
 403         return -EINVAL;
 404 }
 405
 406 static void hugetlb_delete_from_page_cache(struct folio *folio)
 407 {
 408         folio_clear_dirty(folio);
 409         folio_clear_uptodate(folio);
 410         filemap_remove_folio(folio);
 411 }
 412
 413 /*
 414  * Called with i_mmap_rwsem held for inode based vma maps.  This makes
 415  * sure vma (and vm_mm) will not go away.  We also hold the hugetlb fault
 416  * mutex for the page in the mapping.  So, we can not race with page being
 417  * faulted into the vma.
 418  */
 419 static bool hugetlb_vma_maps_page(struct vm_area_struct *vma,
 420                                 unsigned long addr, struct page *page)
 421 {
 422         pte_t *ptep, pte;
 423
 424         ptep = hugetlb_walk(vma, addr, huge_page_size(hstate_vma(vma)));
 425         if (!ptep)
 426                 return false;
 427
 428         pte = huge_ptep_get(ptep);
 429         if (huge_pte_none(pte) || !pte_present(pte))
 430                 return false;
 431
 432         if (pte_page(pte) == page)
 433                 return true;
 434
 435         return false;
 436 }
 437
 438 /*
 439  * Can vma_offset_start/vma_offset_end overflow on 32-bit arches?
 440  * No, because the interval tree returns us only those vmas
 441  * which overlap the truncated area starting at pgoff,
 442  * and no vma on a 32-bit arch can span beyond the 4GB.
 443  */
 444 static unsigned long vma_offset_start(struct vm_area_struct *vma, pgoff_t start)
 445 {
 446         unsigned long offset = 0;
 447
 448         if (vma->vm_pgoff < start)
 449                 offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 450
 451         return vma->vm_start + offset;
 452 }
 453
 454 static unsigned long vma_offset_end(struct vm_area_struct *vma, pgoff_t end)
 455 {
 456         unsigned long t_end;
 457
 458         if (!end)
 459                 return vma->vm_end;
 460
 461         t_end = ((end - vma->vm_pgoff) << PAGE_SHIFT) + vma->vm_start;
 462         if (t_end > vma->vm_end)
 463                 t_end = vma->vm_end;
 464         return t_end;
 465 }
 466
 467 /*
 468  * Called with hugetlb fault mutex held.  Therefore, no more mappings to
 469  * this folio can be created while executing the routine.
 470  */
 471 static void hugetlb_unmap_file_folio(struct hstate *h,
 472                                         struct address_space *mapping,
 473                                         struct folio *folio, pgoff_t index)
 474 {
 475         struct rb_root_cached *root = &mapping->i_mmap;
 476         struct hugetlb_vma_lock *vma_lock;
 477         struct page *page = &folio->page;
 478         struct vm_area_struct *vma;
 479         unsigned long v_start;
 480         unsigned long v_end;
 481         pgoff_t start, end;
 482
 483         start = index * pages_per_huge_page(h);
 484         end = (index + 1) * pages_per_huge_page(h);
 485
 486         i_mmap_lock_write(mapping);
 487 retry:
 488         vma_lock = NULL;
 489         vma_interval_tree_foreach(vma, root, start, end - 1) {
 490                 v_start = vma_offset_start(vma, start);
 491                 v_end = vma_offset_end(vma, end);
 492
 493                 if (!hugetlb_vma_maps_page(vma, v_start, page))
 494                         continue;
 495
 496                 if (!hugetlb_vma_trylock_write(vma)) {
 497                         vma_lock = vma->vm_private_data;
 498                         /*
 499                          * If we can not get vma lock, we need to drop
 500                          * immap_sema and take locks in order.  First,
 501                          * take a ref on the vma_lock structure so that
 502                          * we can be guaranteed it will not go away when
 503                          * dropping immap_sema.
 504                          */
 505                         kref_get(&vma_lock->refs);
 506                         break;
 507                 }
 508
 509                 unmap_hugepage_range(vma, v_start, v_end, NULL,
 510                                      ZAP_FLAG_DROP_MARKER);
 511                 hugetlb_vma_unlock_write(vma);
 512         }
 513
 514         i_mmap_unlock_write(mapping);
 515
 516         if (vma_lock) {
 517                 /*
 518                  * Wait on vma_lock.  We know it is still valid as we have
 519                  * a reference.  We must 'open code' vma locking as we do
 520                  * not know if vma_lock is still attached to vma.
 521                  */
 522                 down_write(&vma_lock->rw_sema);
 523                 i_mmap_lock_write(mapping);
 524
 525                 vma = vma_lock->vma;
 526                 if (!vma) {
 527                         /*
 528                          * If lock is no longer attached to vma, then just
 529                          * unlock, drop our reference and retry looking for
 530                          * other vmas.
 531                          */
 532                         up_write(&vma_lock->rw_sema);
 533                         kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
 534                         goto retry;
 535                 }
 536
 537                 /*
 538                  * vma_lock is still attached to vma.  Check to see if vma
 539                  * still maps page and if so, unmap.
 540                  */
 541                 v_start = vma_offset_start(vma, start);
 542                 v_end = vma_offset_end(vma, end);
 543                 if (hugetlb_vma_maps_page(vma, v_start, page))
 544                         unmap_hugepage_range(vma, v_start, v_end, NULL,
 545                                              ZAP_FLAG_DROP_MARKER);
 546
 547                 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
 548                 hugetlb_vma_unlock_write(vma);
 549
 550                 goto retry;
 551         }
 552 }
 553
 554 static void
 555 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
 556                       zap_flags_t zap_flags)
 557 {
 558         struct vm_area_struct *vma;
 559
 560         /*
 561          * end == 0 indicates that the entire range after start should be
 562          * unmapped.  Note, end is exclusive, whereas the interval tree takes
 563          * an inclusive "last".
 564          */
 565         vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
 566                 unsigned long v_start;
 567                 unsigned long v_end;
 568
 569                 if (!hugetlb_vma_trylock_write(vma))
 570                         continue;
 571
 572                 v_start = vma_offset_start(vma, start);
 573                 v_end = vma_offset_end(vma, end);
 574
 575                 unmap_hugepage_range(vma, v_start, v_end, NULL, zap_flags);
 576
 577                 /*
 578                  * Note that vma lock only exists for shared/non-private
 579                  * vmas.  Therefore, lock is not held when calling
 580                  * unmap_hugepage_range for private vmas.
 581                  */
 582                 hugetlb_vma_unlock_write(vma);
 583         }
 584 }
 585
 586 /*
 587  * Called with hugetlb fault mutex held.
 588  * Returns true if page was actually removed, false otherwise.
 589  */
 590 static bool remove_inode_single_folio(struct hstate *h, struct inode *inode,
 591                                         struct address_space *mapping,
 592                                         struct folio *folio, pgoff_t index,
 593                                         bool truncate_op)
 594 {
 595         bool ret = false;
 596
 597         /*
 598          * If folio is mapped, it was faulted in after being
 599          * unmapped in caller.  Unmap (again) while holding
 600          * the fault mutex.  The mutex will prevent faults
 601          * until we finish removing the folio.
 602          */
 603         if (unlikely(folio_mapped(folio)))
 604                 hugetlb_unmap_file_folio(h, mapping, folio, index);
 605
 606         folio_lock(folio);
 607         /*
 608          * We must remove the folio from page cache before removing
 609          * the region/ reserve map (hugetlb_unreserve_pages).  In
 610          * rare out of memory conditions, removal of the region/reserve
 611          * map could fail.  Correspondingly, the subpool and global
 612          * reserve usage count can need to be adjusted.
 613          */
 614         VM_BUG_ON_FOLIO(folio_test_hugetlb_restore_reserve(folio), folio);
 615         hugetlb_delete_from_page_cache(folio);
 616         ret = true;
 617         if (!truncate_op) {
 618                 if (unlikely(hugetlb_unreserve_pages(inode, index,
 619                                                         index + 1, 1)))
 620                         hugetlb_fix_reserve_counts(inode);
 621         }
 622
 623         folio_unlock(folio);
 624         return ret;
 625 }
 626
 627 /*
 628  * remove_inode_hugepages handles two distinct cases: truncation and hole
 629  * punch.  There are subtle differences in operation for each case.
 630  *
 631  * truncation is indicated by end of range being LLONG_MAX
 632  *      In this case, we first scan the range and release found pages.
 633  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
 634  *      maps and global counts.  Page faults can race with truncation.
 635  *      During faults, hugetlb_no_page() checks i_size before page allocation,
 636  *      and again after obtaining page table lock.  It will 'back out'
 637  *      allocations in the truncated range.
 638  * hole punch is indicated if end is not LLONG_MAX
 639  *      In the hole punch case we scan the range and release found pages.
 640  *      Only when releasing a page is the associated region/reserve map
 641  *      deleted.  The region/reserve map for ranges without associated
 642  *      pages are not modified.  Page faults can race with hole punch.
 643  *      This is indicated if we find a mapped page.
 644  * Note: If the passed end of range value is beyond the end of file, but
 645  * not LLONG_MAX this routine still performs a hole punch operation.
 646  */
 647 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 648                                    loff_t lend)
 649 {
 650         struct hstate *h = hstate_inode(inode);
 651         struct address_space *mapping = &inode->i_data;
 652         const pgoff_t end = lend >> PAGE_SHIFT;
 653         struct folio_batch fbatch;
 654         pgoff_t next, index;
 655         int i, freed = 0;
 656         bool truncate_op = (lend == LLONG_MAX);
 657
 658         folio_batch_init(&fbatch);
 659         next = lstart >> PAGE_SHIFT;
 660         while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
 661                 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
 662                         struct folio *folio = fbatch.folios[i];
 663                         u32 hash = 0;
 664
 665                         index = folio->index >> huge_page_order(h);
 666                         hash = hugetlb_fault_mutex_hash(mapping, index);
 667                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 668
 669                         /*
 670                          * Remove folio that was part of folio_batch.
 671                          */
 672                         if (remove_inode_single_folio(h, inode, mapping, folio,
 673                                                         index, truncate_op))
 674                                 freed++;
 675
 676                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 677                 }
 678                 folio_batch_release(&fbatch);
 679                 cond_resched();
 680         }
 681
 682         if (truncate_op)
 683                 (void)hugetlb_unreserve_pages(inode,
 684                                 lstart >> huge_page_shift(h),
 685                                 LONG_MAX, freed);
 686 }
 687
 688 static void hugetlbfs_evict_inode(struct inode *inode)
 689 {
 690         struct resv_map *resv_map;
 691
 692         remove_inode_hugepages(inode, 0, LLONG_MAX);
 693
 694         /*
 695          * Get the resv_map from the address space embedded in the inode.
 696          * This is the address space which points to any resv_map allocated
 697          * at inode creation time.  If this is a device special inode,
 698          * i_mapping may not point to the original address space.
 699          */
 700         resv_map = (struct resv_map *)(&inode->i_data)->i_private_data;
 701         /* Only regular and link inodes have associated reserve maps */
 702         if (resv_map)
 703                 resv_map_release(&resv_map->refs);
 704         clear_inode(inode);
 705 }
 706
 707 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 708 {
 709         pgoff_t pgoff;
 710         struct address_space *mapping = inode->i_mapping;
 711         struct hstate *h = hstate_inode(inode);
 712
 713         BUG_ON(offset & ~huge_page_mask(h));
 714         pgoff = offset >> PAGE_SHIFT;
 715
 716         i_size_write(inode, offset);
 717         i_mmap_lock_write(mapping);
 718         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 719                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
 720                                       ZAP_FLAG_DROP_MARKER);
 721         i_mmap_unlock_write(mapping);
 722         remove_inode_hugepages(inode, offset, LLONG_MAX);
 723 }
 724
 725 static void hugetlbfs_zero_partial_page(struct hstate *h,
 726                                         struct address_space *mapping,
 727                                         loff_t start,
 728                                         loff_t end)
 729 {
 730         pgoff_t idx = start >> huge_page_shift(h);
 731         struct folio *folio;
 732
 733         folio = filemap_lock_hugetlb_folio(h, mapping, idx);
 734         if (IS_ERR(folio))
 735                 return;
 736
 737         start = start & ~huge_page_mask(h);
 738         end = end & ~huge_page_mask(h);
 739         if (!end)
 740                 end = huge_page_size(h);
 741
 742         folio_zero_segment(folio, (size_t)start, (size_t)end);
 743
 744         folio_unlock(folio);
 745         folio_put(folio);
 746 }
 747
 748 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 749 {
 750         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 751         struct address_space *mapping = inode->i_mapping;
 752         struct hstate *h = hstate_inode(inode);
 753         loff_t hpage_size = huge_page_size(h);
 754         loff_t hole_start, hole_end;
 755
 756         /*
 757          * hole_start and hole_end indicate the full pages within the hole.
 758          */
 759         hole_start = round_up(offset, hpage_size);
 760         hole_end = round_down(offset + len, hpage_size);
 761
 762         inode_lock(inode);
 763
 764         /* protected by i_rwsem */
 765         if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 766                 inode_unlock(inode);
 767                 return -EPERM;
 768         }
 769
 770         i_mmap_lock_write(mapping);
 771
 772         /* If range starts before first full page, zero partial page. */
 773         if (offset < hole_start)
 774                 hugetlbfs_zero_partial_page(h, mapping,
 775                                 offset, min(offset + len, hole_start));
 776
 777         /* Unmap users of full pages in the hole. */
 778         if (hole_end > hole_start) {
 779                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 780                         hugetlb_vmdelete_list(&mapping->i_mmap,
 781                                               hole_start >> PAGE_SHIFT,
 782                                               hole_end >> PAGE_SHIFT, 0);
 783         }
 784
 785         /* If range extends beyond last full page, zero partial page. */
 786         if ((offset + len) > hole_end && (offset + len) > hole_start)
 787                 hugetlbfs_zero_partial_page(h, mapping,
 788                                 hole_end, offset + len);
 789
 790         i_mmap_unlock_write(mapping);
 791
 792         /* Remove full pages from the file. */
 793         if (hole_end > hole_start)
 794                 remove_inode_hugepages(inode, hole_start, hole_end);
 795
 796         inode_unlock(inode);
 797
 798         return 0;
 799 }
 800
 801 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 802                                 loff_t len)
 803 {
 804         struct inode *inode = file_inode(file);
 805         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 806         struct address_space *mapping = inode->i_mapping;
 807         struct hstate *h = hstate_inode(inode);
 808         struct vm_area_struct pseudo_vma;
 809         struct mm_struct *mm = current->mm;
 810         loff_t hpage_size = huge_page_size(h);
 811         unsigned long hpage_shift = huge_page_shift(h);
 812         pgoff_t start, index, end;
 813         int error;
 814         u32 hash;
 815
 816         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 817                 return -EOPNOTSUPP;
 818
 819         if (mode & FALLOC_FL_PUNCH_HOLE)
 820                 return hugetlbfs_punch_hole(inode, offset, len);
 821
 822         /*
 823          * Default preallocate case.
 824          * For this range, start is rounded down and end is rounded up
 825          * as well as being converted to page offsets.
 826          */
 827         start = offset >> hpage_shift;
 828         end = (offset + len + hpage_size - 1) >> hpage_shift;
 829
 830         inode_lock(inode);
 831
 832         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 833         error = inode_newsize_ok(inode, offset + len);
 834         if (error)
 835                 goto out;
 836
 837         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 838                 error = -EPERM;
 839                 goto out;
 840         }
 841
 842         /*
 843          * Initialize a pseudo vma as this is required by the huge page
 844          * allocation routines.
 845          */
 846         vma_init(&pseudo_vma, mm);
 847         vm_flags_init(&pseudo_vma, VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 848         pseudo_vma.vm_file = file;
 849
 850         for (index = start; index < end; index++) {
 851                 /*
 852                  * This is supposed to be the vaddr where the page is being
 853                  * faulted in, but we have no vaddr here.
 854                  */
 855                 struct folio *folio;
 856                 unsigned long addr;
 857
 858                 cond_resched();
 859
 860                 /*
 861                  * fallocate(2) manpage permits EINTR; we may have been
 862                  * interrupted because we are using up too much memory.
 863                  */
 864                 if (signal_pending(current)) {
 865                         error = -EINTR;
 866                         break;
 867                 }
 868
 869                 /* addr is the offset within the file (zero based) */
 870                 addr = index * hpage_size;
 871
 872                 /* mutex taken here, fault path and hole punch */
 873                 hash = hugetlb_fault_mutex_hash(mapping, index);
 874                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 875
 876                 /* See if already present in mapping to avoid alloc/free */
 877                 folio = filemap_get_folio(mapping, index << huge_page_order(h));
 878                 if (!IS_ERR(folio)) {
 879                         folio_put(folio);
 880                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 881                         continue;
 882                 }
 883
 884                 /*
 885                  * Allocate folio without setting the avoid_reserve argument.
 886                  * There certainly are no reserves associated with the
 887                  * pseudo_vma.  However, there could be shared mappings with
 888                  * reserves for the file at the inode level.  If we fallocate
 889                  * folios in these areas, we need to consume the reserves
 890                  * to keep reservation accounting consistent.
 891                  */
 892                 folio = alloc_hugetlb_folio(&pseudo_vma, addr, 0);
 893                 if (IS_ERR(folio)) {
 894                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 895                         error = PTR_ERR(folio);
 896                         goto out;
 897                 }
 898                 clear_huge_page(&folio->page, addr, pages_per_huge_page(h));
 899                 __folio_mark_uptodate(folio);
 900                 error = hugetlb_add_to_page_cache(folio, mapping, index);
 901                 if (unlikely(error)) {
 902                         restore_reserve_on_error(h, &pseudo_vma, addr, folio);
 903                         folio_put(folio);
 904                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 905                         goto out;
 906                 }
 907
 908                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 909
 910                 folio_set_hugetlb_migratable(folio);
 911                 /*
 912                  * folio_unlock because locked by hugetlb_add_to_page_cache()
 913                  * folio_put() due to reference from alloc_hugetlb_folio()
 914                  */
 915                 folio_unlock(folio);
 916                 folio_put(folio);
 917         }
 918
 919         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 920                 i_size_write(inode, offset + len);
 921         inode_set_ctime_current(inode);
 922 out:
 923         inode_unlock(inode);
 924         return error;
 925 }
 926
 927 static int hugetlbfs_setattr(struct mnt_idmap *idmap,
 928                              struct dentry *dentry, struct iattr *attr)
 929 {
 930         struct inode *inode = d_inode(dentry);
 931         struct hstate *h = hstate_inode(inode);
 932         int error;
 933         unsigned int ia_valid = attr->ia_valid;
 934         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 935
 936         error = setattr_prepare(idmap, dentry, attr);
 937         if (error)
 938                 return error;
 939
 940         if (ia_valid & ATTR_SIZE) {
 941                 loff_t oldsize = inode->i_size;
 942                 loff_t newsize = attr->ia_size;
 943
 944                 if (newsize & ~huge_page_mask(h))
 945                         return -EINVAL;
 946                 /* protected by i_rwsem */
 947                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 948                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 949                         return -EPERM;
 950                 hugetlb_vmtruncate(inode, newsize);
 951         }
 952
 953         setattr_copy(idmap, inode, attr);
 954         mark_inode_dirty(inode);
 955         return 0;
 956 }
 957
 958 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 959                                         struct hugetlbfs_fs_context *ctx)
 960 {
 961         struct inode *inode;
 962
 963         inode = new_inode(sb);
 964         if (inode) {
 965                 inode->i_ino = get_next_ino();
 966                 inode->i_mode = S_IFDIR | ctx->mode;
 967                 inode->i_uid = ctx->uid;
 968                 inode->i_gid = ctx->gid;
 969                 simple_inode_init_ts(inode);
 970                 inode->i_op = &hugetlbfs_dir_inode_operations;
 971                 inode->i_fop = &simple_dir_operations;
 972                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 973                 inc_nlink(inode);
 974                 lockdep_annotate_inode_mutex_key(inode);
 975         }
 976         return inode;
 977 }
 978
 979 /*
 980  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 981  * be taken from reclaim -- unlike regular filesystems. This needs an
 982  * annotation because huge_pmd_share() does an allocation under hugetlb's
 983  * i_mmap_rwsem.
 984  */
 985 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 986
 987 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 988                                         struct mnt_idmap *idmap,
 989                                         struct inode *dir,
 990                                         umode_t mode, dev_t dev)
 991 {
 992         struct inode *inode;
 993         struct resv_map *resv_map = NULL;
 994
 995         /*
 996          * Reserve maps are only needed for inodes that can have associated
 997          * page allocations.
 998          */
 999         if (S_ISREG(mode) || S_ISLNK(mode)) {
1000                 resv_map = resv_map_alloc();
1001                 if (!resv_map)
1002                         return NULL;
1003         }
1004
1005         inode = new_inode(sb);
1006         if (inode) {
1007                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
1008
1009                 inode->i_ino = get_next_ino();
1010                 inode_init_owner(idmap, inode, dir, mode);
1011                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
1012                                 &hugetlbfs_i_mmap_rwsem_key);
1013                 inode->i_mapping->a_ops = &hugetlbfs_aops;
1014                 simple_inode_init_ts(inode);
1015                 inode->i_mapping->i_private_data = resv_map;
1016                 info->seals = F_SEAL_SEAL;
1017                 switch (mode & S_IFMT) {
1018                 default:
1019                         init_special_inode(inode, mode, dev);
1020                         break;
1021                 case S_IFREG:
1022                         inode->i_op = &hugetlbfs_inode_operations;
1023                         inode->i_fop = &hugetlbfs_file_operations;
1024                         break;
1025                 case S_IFDIR:
1026                         inode->i_op = &hugetlbfs_dir_inode_operations;
1027                         inode->i_fop = &simple_dir_operations;
1028
1029                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
1030                         inc_nlink(inode);
1031                         break;
1032                 case S_IFLNK:
1033                         inode->i_op = &page_symlink_inode_operations;
1034                         inode_nohighmem(inode);
1035                         break;
1036                 }
1037                 lockdep_annotate_inode_mutex_key(inode);
1038         } else {
1039                 if (resv_map)
1040                         kref_put(&resv_map->refs, resv_map_release);
1041         }
1042
1043         return inode;
1044 }
1045
1046 /*
1047  * File creation. Allocate an inode, and we're done..
1048  */
1049 static int hugetlbfs_mknod(struct mnt_idmap *idmap, struct inode *dir,
1050                            struct dentry *dentry, umode_t mode, dev_t dev)
1051 {
1052         struct inode *inode;
1053
1054         inode = hugetlbfs_get_inode(dir->i_sb, idmap, dir, mode, dev);
1055         if (!inode)
1056                 return -ENOSPC;
1057         inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1058         d_instantiate(dentry, inode);
1059         dget(dentry);/* Extra count - pin the dentry in core */
1060         return 0;
1061 }
1062
1063 static int hugetlbfs_mkdir(struct mnt_idmap *idmap, struct inode *dir,
1064                            struct dentry *dentry, umode_t mode)
1065 {
1066         int retval = hugetlbfs_mknod(idmap, dir, dentry,
1067                                      mode | S_IFDIR, 0);
1068         if (!retval)
1069                 inc_nlink(dir);
1070         return retval;
1071 }
1072
1073 static int hugetlbfs_create(struct mnt_idmap *idmap,
1074                             struct inode *dir, struct dentry *dentry,
1075                             umode_t mode, bool excl)
1076 {
1077         return hugetlbfs_mknod(idmap, dir, dentry, mode | S_IFREG, 0);
1078 }
1079
1080 static int hugetlbfs_tmpfile(struct mnt_idmap *idmap,
1081                              struct inode *dir, struct file *file,
1082                              umode_t mode)
1083 {
1084         struct inode *inode;
1085
1086         inode = hugetlbfs_get_inode(dir->i_sb, idmap, dir, mode | S_IFREG, 0);
1087         if (!inode)
1088                 return -ENOSPC;
1089         inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1090         d_tmpfile(file, inode);
1091         return finish_open_simple(file, 0);
1092 }
1093
1094 static int hugetlbfs_symlink(struct mnt_idmap *idmap,
1095                              struct inode *dir, struct dentry *dentry,
1096                              const char *symname)
1097 {
1098         const umode_t mode = S_IFLNK|S_IRWXUGO;
1099         struct inode *inode;
1100         int error = -ENOSPC;
1101
1102         inode = hugetlbfs_get_inode(dir->i_sb, idmap, dir, mode, 0);
1103         if (inode) {
1104                 int l = strlen(symname)+1;
1105                 error = page_symlink(inode, symname, l);
1106                 if (!error) {
1107                         d_instantiate(dentry, inode);
1108                         dget(dentry);
1109                 } else
1110                         iput(inode);
1111         }
1112         inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1113
1114         return error;
1115 }
1116
1117 #ifdef CONFIG_MIGRATION
1118 static int hugetlbfs_migrate_folio(struct address_space *mapping,
1119                                 struct folio *dst, struct folio *src,
1120                                 enum migrate_mode mode)
1121 {
1122         int rc;
1123
1124         rc = migrate_huge_page_move_mapping(mapping, dst, src);
1125         if (rc != MIGRATEPAGE_SUCCESS)
1126                 return rc;
1127
1128         if (hugetlb_folio_subpool(src)) {
1129                 hugetlb_set_folio_subpool(dst,
1130                                         hugetlb_folio_subpool(src));
1131                 hugetlb_set_folio_subpool(src, NULL);
1132         }
1133
1134         if (mode != MIGRATE_SYNC_NO_COPY)
1135                 folio_migrate_copy(dst, src);
1136         else
1137                 folio_migrate_flags(dst, src);
1138
1139         return MIGRATEPAGE_SUCCESS;
1140 }
1141 #else
1142 #define hugetlbfs_migrate_folio NULL
1143 #endif
1144
1145 static int hugetlbfs_error_remove_folio(struct address_space *mapping,
1146                                 struct folio *folio)
1147 {
1148         return 0;
1149 }
1150
1151 /*
1152  * Display the mount options in /proc/mounts.
1153  */
1154 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1155 {
1156         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1157         struct hugepage_subpool *spool = sbinfo->spool;
1158         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1159         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1160         char mod;
1161
1162         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1163                 seq_printf(m, ",uid=%u",
1164                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1165         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1166                 seq_printf(m, ",gid=%u",
1167                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1168         if (sbinfo->mode != 0755)
1169                 seq_printf(m, ",mode=%o", sbinfo->mode);
1170         if (sbinfo->max_inodes != -1)
1171                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1172
1173         hpage_size /= 1024;
1174         mod = 'K';
1175         if (hpage_size >= 1024) {
1176                 hpage_size /= 1024;
1177                 mod = 'M';
1178         }
1179         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1180         if (spool) {
1181                 if (spool->max_hpages != -1)
1182                         seq_printf(m, ",size=%llu",
1183                                    (unsigned long long)spool->max_hpages << hpage_shift);
1184                 if (spool->min_hpages != -1)
1185                         seq_printf(m, ",min_size=%llu",
1186                                    (unsigned long long)spool->min_hpages << hpage_shift);
1187         }
1188         return 0;
1189 }
1190
1191 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1192 {
1193         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1194         struct hstate *h = hstate_inode(d_inode(dentry));
1195         u64 id = huge_encode_dev(dentry->d_sb->s_dev);
1196
1197         buf->f_fsid = u64_to_fsid(id);
1198         buf->f_type = HUGETLBFS_MAGIC;
1199         buf->f_bsize = huge_page_size(h);
1200         if (sbinfo) {
1201                 spin_lock(&sbinfo->stat_lock);
1202                 /* If no limits set, just report 0 or -1 for max/free/used
1203                  * blocks, like simple_statfs() */
1204                 if (sbinfo->spool) {
1205                         long free_pages;
1206
1207                         spin_lock_irq(&sbinfo->spool->lock);
1208                         buf->f_blocks = sbinfo->spool->max_hpages;
1209                         free_pages = sbinfo->spool->max_hpages
1210                                 - sbinfo->spool->used_hpages;
1211                         buf->f_bavail = buf->f_bfree = free_pages;
1212                         spin_unlock_irq(&sbinfo->spool->lock);
1213                         buf->f_files = sbinfo->max_inodes;
1214                         buf->f_ffree = sbinfo->free_inodes;
1215                 }
1216                 spin_unlock(&sbinfo->stat_lock);
1217         }
1218         buf->f_namelen = NAME_MAX;
1219         return 0;
1220 }
1221
1222 static void hugetlbfs_put_super(struct super_block *sb)
1223 {
1224         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1225
1226         if (sbi) {
1227                 sb->s_fs_info = NULL;
1228
1229                 if (sbi->spool)
1230                         hugepage_put_subpool(sbi->spool);
1231
1232                 kfree(sbi);
1233         }
1234 }
1235
1236 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1237 {
1238         if (sbinfo->free_inodes >= 0) {
1239                 spin_lock(&sbinfo->stat_lock);
1240                 if (unlikely(!sbinfo->free_inodes)) {
1241                         spin_unlock(&sbinfo->stat_lock);
1242                         return 0;
1243                 }
1244                 sbinfo->free_inodes--;
1245                 spin_unlock(&sbinfo->stat_lock);
1246         }
1247
1248         return 1;
1249 }
1250
1251 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1252 {
1253         if (sbinfo->free_inodes >= 0) {
1254                 spin_lock(&sbinfo->stat_lock);
1255                 sbinfo->free_inodes++;
1256                 spin_unlock(&sbinfo->stat_lock);
1257         }
1258 }
1259
1260
1261 static struct kmem_cache *hugetlbfs_inode_cachep;
1262
1263 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1264 {
1265         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1266         struct hugetlbfs_inode_info *p;
1267
1268         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1269                 return NULL;
1270         p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1271         if (unlikely(!p)) {
1272                 hugetlbfs_inc_free_inodes(sbinfo);
1273                 return NULL;
1274         }
1275         return &p->vfs_inode;
1276 }
1277
1278 static void hugetlbfs_free_inode(struct inode *inode)
1279 {
1280         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1281 }
1282
1283 static void hugetlbfs_destroy_inode(struct inode *inode)
1284 {
1285         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1286 }
1287
1288 static const struct address_space_operations hugetlbfs_aops = {
1289         .write_begin    = hugetlbfs_write_begin,
1290         .write_end      = hugetlbfs_write_end,
1291         .dirty_folio    = noop_dirty_folio,
1292         .migrate_folio  = hugetlbfs_migrate_folio,
1293         .error_remove_folio     = hugetlbfs_error_remove_folio,
1294 };
1295
1296
1297 static void init_once(void *foo)
1298 {
1299         struct hugetlbfs_inode_info *ei = foo;
1300
1301         inode_init_once(&ei->vfs_inode);
1302 }
1303
1304 const struct file_operations hugetlbfs_file_operations = {
1305         .read_iter              = hugetlbfs_read_iter,
1306         .mmap                   = hugetlbfs_file_mmap,
1307         .fsync                  = noop_fsync,
1308         .get_unmapped_area      = hugetlb_get_unmapped_area,
1309         .llseek                 = default_llseek,
1310         .fallocate              = hugetlbfs_fallocate,
1311 };
1312
1313 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1314         .create         = hugetlbfs_create,
1315         .lookup         = simple_lookup,
1316         .link           = simple_link,
1317         .unlink         = simple_unlink,
1318         .symlink        = hugetlbfs_symlink,
1319         .mkdir          = hugetlbfs_mkdir,
1320         .rmdir          = simple_rmdir,
1321         .mknod          = hugetlbfs_mknod,
1322         .rename         = simple_rename,
1323         .setattr        = hugetlbfs_setattr,
1324         .tmpfile        = hugetlbfs_tmpfile,
1325 };
1326
1327 static const struct inode_operations hugetlbfs_inode_operations = {
1328         .setattr        = hugetlbfs_setattr,
1329 };
1330
1331 static const struct super_operations hugetlbfs_ops = {
1332         .alloc_inode    = hugetlbfs_alloc_inode,
1333         .free_inode     = hugetlbfs_free_inode,
1334         .destroy_inode  = hugetlbfs_destroy_inode,
1335         .evict_inode    = hugetlbfs_evict_inode,
1336         .statfs         = hugetlbfs_statfs,
1337         .put_super      = hugetlbfs_put_super,
1338         .show_options   = hugetlbfs_show_options,
1339 };
1340
1341 /*
1342  * Convert size option passed from command line to number of huge pages
1343  * in the pool specified by hstate.  Size option could be in bytes
1344  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1345  */
1346 static long
1347 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1348                          enum hugetlbfs_size_type val_type)
1349 {
1350         if (val_type == NO_SIZE)
1351                 return -1;
1352
1353         if (val_type == SIZE_PERCENT) {
1354                 size_opt <<= huge_page_shift(h);
1355                 size_opt *= h->max_huge_pages;
1356                 do_div(size_opt, 100);
1357         }
1358
1359         size_opt >>= huge_page_shift(h);
1360         return size_opt;
1361 }
1362
1363 /*
1364  * Parse one mount parameter.
1365  */
1366 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1367 {
1368         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1369         struct fs_parse_result result;
1370         struct hstate *h;
1371         char *rest;
1372         unsigned long ps;
1373         int opt;
1374
1375         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1376         if (opt < 0)
1377                 return opt;
1378
1379         switch (opt) {
1380         case Opt_uid:
1381                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1382                 if (!uid_valid(ctx->uid))
1383                         goto bad_val;
1384                 return 0;
1385
1386         case Opt_gid:
1387                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1388                 if (!gid_valid(ctx->gid))
1389                         goto bad_val;
1390                 return 0;
1391
1392         case Opt_mode:
1393                 ctx->mode = result.uint_32 & 01777U;
1394                 return 0;
1395
1396         case Opt_size:
1397                 /* memparse() will accept a K/M/G without a digit */
1398                 if (!param->string || !isdigit(param->string[0]))
1399                         goto bad_val;
1400                 ctx->max_size_opt = memparse(param->string, &rest);
1401                 ctx->max_val_type = SIZE_STD;
1402                 if (*rest == '%')
1403                         ctx->max_val_type = SIZE_PERCENT;
1404                 return 0;
1405
1406         case Opt_nr_inodes:
1407                 /* memparse() will accept a K/M/G without a digit */
1408                 if (!param->string || !isdigit(param->string[0]))
1409                         goto bad_val;
1410                 ctx->nr_inodes = memparse(param->string, &rest);
1411                 return 0;
1412
1413         case Opt_pagesize:
1414                 ps = memparse(param->string, &rest);
1415                 h = size_to_hstate(ps);
1416                 if (!h) {
1417                         pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1418                         return -EINVAL;
1419                 }
1420                 ctx->hstate = h;
1421                 return 0;
1422
1423         case Opt_min_size:
1424                 /* memparse() will accept a K/M/G without a digit */
1425                 if (!param->string || !isdigit(param->string[0]))
1426                         goto bad_val;
1427                 ctx->min_size_opt = memparse(param->string, &rest);
1428                 ctx->min_val_type = SIZE_STD;
1429                 if (*rest == '%')
1430                         ctx->min_val_type = SIZE_PERCENT;
1431                 return 0;
1432
1433         default:
1434                 return -EINVAL;
1435         }
1436
1437 bad_val:
1438         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1439                       param->string, param->key);
1440 }
1441
1442 /*
1443  * Validate the parsed options.
1444  */
1445 static int hugetlbfs_validate(struct fs_context *fc)
1446 {
1447         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1448
1449         /*
1450          * Use huge page pool size (in hstate) to convert the size
1451          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1452          */
1453         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1454                                                    ctx->max_size_opt,
1455                                                    ctx->max_val_type);
1456         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1457                                                    ctx->min_size_opt,
1458                                                    ctx->min_val_type);
1459
1460         /*
1461          * If max_size was specified, then min_size must be smaller
1462          */
1463         if (ctx->max_val_type > NO_SIZE &&
1464             ctx->min_hpages > ctx->max_hpages) {
1465                 pr_err("Minimum size can not be greater than maximum size\n");
1466                 return -EINVAL;
1467         }
1468
1469         return 0;
1470 }
1471
1472 static int
1473 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1474 {
1475         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1476         struct hugetlbfs_sb_info *sbinfo;
1477
1478         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1479         if (!sbinfo)
1480                 return -ENOMEM;
1481         sb->s_fs_info = sbinfo;
1482         spin_lock_init(&sbinfo->stat_lock);
1483         sbinfo->hstate          = ctx->hstate;
1484         sbinfo->max_inodes      = ctx->nr_inodes;
1485         sbinfo->free_inodes     = ctx->nr_inodes;
1486         sbinfo->spool           = NULL;
1487         sbinfo->uid             = ctx->uid;
1488         sbinfo->gid             = ctx->gid;
1489         sbinfo->mode            = ctx->mode;
1490
1491         /*
1492          * Allocate and initialize subpool if maximum or minimum size is
1493          * specified.  Any needed reservations (for minimum size) are taken
1494          * when the subpool is created.
1495          */
1496         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1497                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1498                                                      ctx->max_hpages,
1499                                                      ctx->min_hpages);
1500                 if (!sbinfo->spool)
1501                         goto out_free;
1502         }
1503         sb->s_maxbytes = MAX_LFS_FILESIZE;
1504         sb->s_blocksize = huge_page_size(ctx->hstate);
1505         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1506         sb->s_magic = HUGETLBFS_MAGIC;
1507         sb->s_op = &hugetlbfs_ops;
1508         sb->s_time_gran = 1;
1509
1510         /*
1511          * Due to the special and limited functionality of hugetlbfs, it does
1512          * not work well as a stacking filesystem.
1513          */
1514         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1515         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1516         if (!sb->s_root)
1517                 goto out_free;
1518         return 0;
1519 out_free:
1520         kfree(sbinfo->spool);
1521         kfree(sbinfo);
1522         return -ENOMEM;
1523 }
1524
1525 static int hugetlbfs_get_tree(struct fs_context *fc)
1526 {
1527         int err = hugetlbfs_validate(fc);
1528         if (err)
1529                 return err;
1530         return get_tree_nodev(fc, hugetlbfs_fill_super);
1531 }
1532
1533 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1534 {
1535         kfree(fc->fs_private);
1536 }
1537
1538 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1539         .free           = hugetlbfs_fs_context_free,
1540         .parse_param    = hugetlbfs_parse_param,
1541         .get_tree       = hugetlbfs_get_tree,
1542 };
1543
1544 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1545 {
1546         struct hugetlbfs_fs_context *ctx;
1547
1548         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1549         if (!ctx)
1550                 return -ENOMEM;
1551
1552         ctx->max_hpages = -1; /* No limit on size by default */
1553         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1554         ctx->uid        = current_fsuid();
1555         ctx->gid        = current_fsgid();
1556         ctx->mode       = 0755;
1557         ctx->hstate     = &default_hstate;
1558         ctx->min_hpages = -1; /* No default minimum size */
1559         ctx->max_val_type = NO_SIZE;
1560         ctx->min_val_type = NO_SIZE;
1561         fc->fs_private = ctx;
1562         fc->ops = &hugetlbfs_fs_context_ops;
1563         return 0;
1564 }
1565
1566 static struct file_system_type hugetlbfs_fs_type = {
1567         .name                   = "hugetlbfs",
1568         .init_fs_context        = hugetlbfs_init_fs_context,
1569         .parameters             = hugetlb_fs_parameters,
1570         .kill_sb                = kill_litter_super,
1571         .fs_flags               = FS_ALLOW_IDMAP,
1572 };
1573
1574 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1575
1576 static int can_do_hugetlb_shm(void)
1577 {
1578         kgid_t shm_group;
1579         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1580         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1581 }
1582
1583 static int get_hstate_idx(int page_size_log)
1584 {
1585         struct hstate *h = hstate_sizelog(page_size_log);
1586
1587         if (!h)
1588                 return -1;
1589         return hstate_index(h);
1590 }
1591
1592 /*
1593  * Note that size should be aligned to proper hugepage size in caller side,
1594  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1595  */
1596 struct file *hugetlb_file_setup(const char *name, size_t size,
1597                                 vm_flags_t acctflag, int creat_flags,
1598                                 int page_size_log)
1599 {
1600         struct inode *inode;
1601         struct vfsmount *mnt;
1602         int hstate_idx;
1603         struct file *file;
1604
1605         hstate_idx = get_hstate_idx(page_size_log);
1606         if (hstate_idx < 0)
1607                 return ERR_PTR(-ENODEV);
1608
1609         mnt = hugetlbfs_vfsmount[hstate_idx];
1610         if (!mnt)
1611                 return ERR_PTR(-ENOENT);
1612
1613         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1614                 struct ucounts *ucounts = current_ucounts();
1615
1616                 if (user_shm_lock(size, ucounts)) {
1617                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1618                                 current->comm, current->pid);
1619                         user_shm_unlock(size, ucounts);
1620                 }
1621                 return ERR_PTR(-EPERM);
1622         }
1623
1624         file = ERR_PTR(-ENOSPC);
1625         /* hugetlbfs_vfsmount[] mounts do not use idmapped mounts.  */
1626         inode = hugetlbfs_get_inode(mnt->mnt_sb, &nop_mnt_idmap, NULL,
1627                                     S_IFREG | S_IRWXUGO, 0);
1628         if (!inode)
1629                 goto out;
1630         if (creat_flags == HUGETLB_SHMFS_INODE)
1631                 inode->i_flags |= S_PRIVATE;
1632
1633         inode->i_size = size;
1634         clear_nlink(inode);
1635
1636         if (!hugetlb_reserve_pages(inode, 0,
1637                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1638                         acctflag))
1639                 file = ERR_PTR(-ENOMEM);
1640         else
1641                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1642                                         &hugetlbfs_file_operations);
1643         if (!IS_ERR(file))
1644                 return file;
1645
1646         iput(inode);
1647 out:
1648         return file;
1649 }
1650
1651 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1652 {
1653         struct fs_context *fc;
1654         struct vfsmount *mnt;
1655
1656         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1657         if (IS_ERR(fc)) {
1658                 mnt = ERR_CAST(fc);
1659         } else {
1660                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1661                 ctx->hstate = h;
1662                 mnt = fc_mount(fc);
1663                 put_fs_context(fc);
1664         }
1665         if (IS_ERR(mnt))
1666                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1667                        huge_page_size(h) / SZ_1K);
1668         return mnt;
1669 }
1670
1671 static int __init init_hugetlbfs_fs(void)
1672 {
1673         struct vfsmount *mnt;
1674         struct hstate *h;
1675         int error;
1676         int i;
1677
1678         if (!hugepages_supported()) {
1679                 pr_info("disabling because there are no supported hugepage sizes\n");
1680                 return -ENOTSUPP;
1681         }
1682
1683         error = -ENOMEM;
1684         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1685                                         sizeof(struct hugetlbfs_inode_info),
1686                                         0, SLAB_ACCOUNT, init_once);
1687         if (hugetlbfs_inode_cachep == NULL)
1688                 goto out;
1689
1690         error = register_filesystem(&hugetlbfs_fs_type);
1691         if (error)
1692                 goto out_free;
1693
1694         /* default hstate mount is required */
1695         mnt = mount_one_hugetlbfs(&default_hstate);
1696         if (IS_ERR(mnt)) {
1697                 error = PTR_ERR(mnt);
1698                 goto out_unreg;
1699         }
1700         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1701
1702         /* other hstates are optional */
1703         i = 0;
1704         for_each_hstate(h) {
1705                 if (i == default_hstate_idx) {
1706                         i++;
1707                         continue;
1708                 }
1709
1710                 mnt = mount_one_hugetlbfs(h);
1711                 if (IS_ERR(mnt))
1712                         hugetlbfs_vfsmount[i] = NULL;
1713                 else
1714                         hugetlbfs_vfsmount[i] = mnt;
1715                 i++;
1716         }
1717
1718         return 0;
1719
1720  out_unreg:
1721         (void)unregister_filesystem(&hugetlbfs_fs_type);
1722  out_free:
1723         kmem_cache_destroy(hugetlbfs_inode_cachep);
1724  out:
1725         return error;
1726 }
1727 fs_initcall(init_hugetlbfs_fs)