hugetlbfs/inode.c

   1 /*
   2  * hugetlbpage-backed filesystem.  Based on ramfs.
   3  *
   4  * Nadia Yvette Chambers, 2002
   5  *
   6  * Copyright (C) 2002 Linus Torvalds.
   7  * License: GPL
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/thread_info.h>
  13 #include <asm/current.h>
  14 #include <linux/falloc.h>
  15 #include <linux/fs.h>
  16 #include <linux/mount.h>
  17 #include <linux/file.h>
  18 #include <linux/kernel.h>
  19 #include <linux/writeback.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/highmem.h>
  22 #include <linux/init.h>
  23 #include <linux/string.h>
  24 #include <linux/capability.h>
  25 #include <linux/ctype.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/hugetlb.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/fs_parser.h>
  30 #include <linux/mman.h>
  31 #include <linux/slab.h>
  32 #include <linux/dnotify.h>
  33 #include <linux/statfs.h>
  34 #include <linux/security.h>
  35 #include <linux/magic.h>
  36 #include <linux/migrate.h>
  37 #include <linux/uio.h>
  38
  39 #include <linux/uaccess.h>
  40 #include <linux/sched/mm.h>
  41
  42 static const struct address_space_operations hugetlbfs_aops;
  43 const struct file_operations hugetlbfs_file_operations;
  44 static const struct inode_operations hugetlbfs_dir_inode_operations;
  45 static const struct inode_operations hugetlbfs_inode_operations;
  46
  47 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
  48
  49 struct hugetlbfs_fs_context {
  50         struct hstate           *hstate;
  51         unsigned long long      max_size_opt;
  52         unsigned long long      min_size_opt;
  53         long                    max_hpages;
  54         long                    nr_inodes;
  55         long                    min_hpages;
  56         enum hugetlbfs_size_type max_val_type;
  57         enum hugetlbfs_size_type min_val_type;
  58         kuid_t                  uid;
  59         kgid_t                  gid;
  60         umode_t                 mode;
  61 };
  62
  63 int sysctl_hugetlb_shm_group;
  64
  65 enum hugetlb_param {
  66         Opt_gid,
  67         Opt_min_size,
  68         Opt_mode,
  69         Opt_nr_inodes,
  70         Opt_pagesize,
  71         Opt_size,
  72         Opt_uid,
  73 };
  74
  75 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
  76         fsparam_u32   ("gid",           Opt_gid),
  77         fsparam_string("min_size",      Opt_min_size),
  78         fsparam_u32oct("mode",          Opt_mode),
  79         fsparam_string("nr_inodes",     Opt_nr_inodes),
  80         fsparam_string("pagesize",      Opt_pagesize),
  81         fsparam_string("size",          Opt_size),
  82         fsparam_u32   ("uid",           Opt_uid),
  83         {}
  84 };
  85
  86 #ifdef CONFIG_NUMA
  87 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  88                                         struct inode *inode, pgoff_t index)
  89 {
  90         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  91                                                         index);
  92 }
  93
  94 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  95 {
  96         mpol_cond_put(vma->vm_policy);
  97 }
  98 #else
  99 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
 100                                         struct inode *inode, pgoff_t index)
 101 {
 102 }
 103
 104 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
 105 {
 106 }
 107 #endif
 108
 109 /*
 110  * Mask used when checking the page offset value passed in via system
 111  * calls.  This value will be converted to a loff_t which is signed.
 112  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 113  * value.  The extra bit (- 1 in the shift value) is to take the sign
 114  * bit into account.
 115  */
 116 #define PGOFF_LOFFT_MAX \
 117         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 118
 119 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 120 {
 121         struct inode *inode = file_inode(file);
 122         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 123         loff_t len, vma_len;
 124         int ret;
 125         struct hstate *h = hstate_file(file);
 126         vm_flags_t vm_flags;
 127
 128         /*
 129          * vma address alignment (but not the pgoff alignment) has
 130          * already been checked by prepare_hugepage_range.  If you add
 131          * any error returns here, do so after setting VM_HUGETLB, so
 132          * is_vm_hugetlb_page tests below unmap_region go the right
 133          * way when do_mmap unwinds (may be important on powerpc
 134          * and ia64).
 135          */
 136         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 137         vma->vm_ops = &hugetlb_vm_ops;
 138
 139         ret = seal_check_future_write(info->seals, vma);
 140         if (ret)
 141                 return ret;
 142
 143         /*
 144          * page based offset in vm_pgoff could be sufficiently large to
 145          * overflow a loff_t when converted to byte offset.  This can
 146          * only happen on architectures where sizeof(loff_t) ==
 147          * sizeof(unsigned long).  So, only check in those instances.
 148          */
 149         if (sizeof(unsigned long) == sizeof(loff_t)) {
 150                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 151                         return -EINVAL;
 152         }
 153
 154         /* must be huge page aligned */
 155         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 156                 return -EINVAL;
 157
 158         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 159         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 160         /* check for overflow */
 161         if (len < vma_len)
 162                 return -EINVAL;
 163
 164         inode_lock(inode);
 165         file_accessed(file);
 166
 167         ret = -ENOMEM;
 168
 169         vm_flags = vma->vm_flags;
 170         /*
 171          * for SHM_HUGETLB, the pages are reserved in the shmget() call so skip
 172          * reserving here. Note: only for SHM hugetlbfs file, the inode
 173          * flag S_PRIVATE is set.
 174          */
 175         if (inode->i_flags & S_PRIVATE)
 176                 vm_flags |= VM_NORESERVE;
 177
 178         if (!hugetlb_reserve_pages(inode,
 179                                 vma->vm_pgoff >> huge_page_order(h),
 180                                 len >> huge_page_shift(h), vma,
 181                                 vm_flags))
 182                 goto out;
 183
 184         ret = 0;
 185         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 186                 i_size_write(inode, len);
 187 out:
 188         inode_unlock(inode);
 189
 190         return ret;
 191 }
 192
 193 /*
 194  * Called under mmap_write_lock(mm).
 195  */
 196
 197 static unsigned long
 198 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
 199                 unsigned long len, unsigned long pgoff, unsigned long flags)
 200 {
 201         struct hstate *h = hstate_file(file);
 202         struct vm_unmapped_area_info info;
 203
 204         info.flags = 0;
 205         info.length = len;
 206         info.low_limit = current->mm->mmap_base;
 207         info.high_limit = arch_get_mmap_end(addr, len, flags);
 208         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 209         info.align_offset = 0;
 210         return vm_unmapped_area(&info);
 211 }
 212
 213 static unsigned long
 214 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
 215                 unsigned long len, unsigned long pgoff, unsigned long flags)
 216 {
 217         struct hstate *h = hstate_file(file);
 218         struct vm_unmapped_area_info info;
 219
 220         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 221         info.length = len;
 222         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
 223         info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
 224         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 225         info.align_offset = 0;
 226         addr = vm_unmapped_area(&info);
 227
 228         /*
 229          * A failed mmap() very likely causes application failure,
 230          * so fall back to the bottom-up function here. This scenario
 231          * can happen with large stack limits and large mmap()
 232          * allocations.
 233          */
 234         if (unlikely(offset_in_page(addr))) {
 235                 VM_BUG_ON(addr != -ENOMEM);
 236                 info.flags = 0;
 237                 info.low_limit = current->mm->mmap_base;
 238                 info.high_limit = arch_get_mmap_end(addr, len, flags);
 239                 addr = vm_unmapped_area(&info);
 240         }
 241
 242         return addr;
 243 }
 244
 245 unsigned long
 246 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 247                                   unsigned long len, unsigned long pgoff,
 248                                   unsigned long flags)
 249 {
 250         struct mm_struct *mm = current->mm;
 251         struct vm_area_struct *vma;
 252         struct hstate *h = hstate_file(file);
 253         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
 254
 255         if (len & ~huge_page_mask(h))
 256                 return -EINVAL;
 257         if (len > TASK_SIZE)
 258                 return -ENOMEM;
 259
 260         if (flags & MAP_FIXED) {
 261                 if (prepare_hugepage_range(file, addr, len))
 262                         return -EINVAL;
 263                 return addr;
 264         }
 265
 266         if (addr) {
 267                 addr = ALIGN(addr, huge_page_size(h));
 268                 vma = find_vma(mm, addr);
 269                 if (mmap_end - len >= addr &&
 270                     (!vma || addr + len <= vm_start_gap(vma)))
 271                         return addr;
 272         }
 273
 274         /*
 275          * Use mm->get_unmapped_area value as a hint to use topdown routine.
 276          * If architectures have special needs, they should define their own
 277          * version of hugetlb_get_unmapped_area.
 278          */
 279         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
 280                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 281                                 pgoff, flags);
 282         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 283                         pgoff, flags);
 284 }
 285
 286 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 287 static unsigned long
 288 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 289                           unsigned long len, unsigned long pgoff,
 290                           unsigned long flags)
 291 {
 292         return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
 293 }
 294 #endif
 295
 296 /*
 297  * Support for read() - Find the page attached to f_mapping and copy out the
 298  * data. This provides functionality similar to filemap_read().
 299  */
 300 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 301 {
 302         struct file *file = iocb->ki_filp;
 303         struct hstate *h = hstate_file(file);
 304         struct address_space *mapping = file->f_mapping;
 305         struct inode *inode = mapping->host;
 306         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 307         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 308         unsigned long end_index;
 309         loff_t isize;
 310         ssize_t retval = 0;
 311
 312         while (iov_iter_count(to)) {
 313                 struct page *page;
 314                 size_t nr, copied;
 315
 316                 /* nr is the maximum number of bytes to copy from this page */
 317                 nr = huge_page_size(h);
 318                 isize = i_size_read(inode);
 319                 if (!isize)
 320                         break;
 321                 end_index = (isize - 1) >> huge_page_shift(h);
 322                 if (index > end_index)
 323                         break;
 324                 if (index == end_index) {
 325                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 326                         if (nr <= offset)
 327                                 break;
 328                 }
 329                 nr = nr - offset;
 330
 331                 /* Find the page */
 332                 page = find_lock_page(mapping, index);
 333                 if (unlikely(page == NULL)) {
 334                         /*
 335                          * We have a HOLE, zero out the user-buffer for the
 336                          * length of the hole or request.
 337                          */
 338                         copied = iov_iter_zero(nr, to);
 339                 } else {
 340                         unlock_page(page);
 341
 342                         if (PageHWPoison(page)) {
 343                                 put_page(page);
 344                                 retval = -EIO;
 345                                 break;
 346                         }
 347
 348                         /*
 349                          * We have the page, copy it to user space buffer.
 350                          */
 351                         copied = copy_page_to_iter(page, offset, nr, to);
 352                         put_page(page);
 353                 }
 354                 offset += copied;
 355                 retval += copied;
 356                 if (copied != nr && iov_iter_count(to)) {
 357                         if (!retval)
 358                                 retval = -EFAULT;
 359                         break;
 360                 }
 361                 index += offset >> huge_page_shift(h);
 362                 offset &= ~huge_page_mask(h);
 363         }
 364         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 365         return retval;
 366 }
 367
 368 static int hugetlbfs_write_begin(struct file *file,
 369                         struct address_space *mapping,
 370                         loff_t pos, unsigned len,
 371                         struct page **pagep, void **fsdata)
 372 {
 373         return -EINVAL;
 374 }
 375
 376 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 377                         loff_t pos, unsigned len, unsigned copied,
 378                         struct page *page, void *fsdata)
 379 {
 380         BUG();
 381         return -EINVAL;
 382 }
 383
 384 static void hugetlb_delete_from_page_cache(struct page *page)
 385 {
 386         ClearPageDirty(page);
 387         ClearPageUptodate(page);
 388         delete_from_page_cache(page);
 389 }
 390
 391 /*
 392  * Called with i_mmap_rwsem held for inode based vma maps.  This makes
 393  * sure vma (and vm_mm) will not go away.  We also hold the hugetlb fault
 394  * mutex for the page in the mapping.  So, we can not race with page being
 395  * faulted into the vma.
 396  */
 397 static bool hugetlb_vma_maps_page(struct vm_area_struct *vma,
 398                                 unsigned long addr, struct page *page)
 399 {
 400         pte_t *ptep, pte;
 401
 402         ptep = huge_pte_offset(vma->vm_mm, addr,
 403                         huge_page_size(hstate_vma(vma)));
 404
 405         if (!ptep)
 406                 return false;
 407
 408         pte = huge_ptep_get(ptep);
 409         if (huge_pte_none(pte) || !pte_present(pte))
 410                 return false;
 411
 412         if (pte_page(pte) == page)
 413                 return true;
 414
 415         return false;
 416 }
 417
 418 /*
 419  * Can vma_offset_start/vma_offset_end overflow on 32-bit arches?
 420  * No, because the interval tree returns us only those vmas
 421  * which overlap the truncated area starting at pgoff,
 422  * and no vma on a 32-bit arch can span beyond the 4GB.
 423  */
 424 static unsigned long vma_offset_start(struct vm_area_struct *vma, pgoff_t start)
 425 {
 426         if (vma->vm_pgoff < start)
 427                 return (start - vma->vm_pgoff) << PAGE_SHIFT;
 428         else
 429                 return 0;
 430 }
 431
 432 static unsigned long vma_offset_end(struct vm_area_struct *vma, pgoff_t end)
 433 {
 434         unsigned long t_end;
 435
 436         if (!end)
 437                 return vma->vm_end;
 438
 439         t_end = ((end - vma->vm_pgoff) << PAGE_SHIFT) + vma->vm_start;
 440         if (t_end > vma->vm_end)
 441                 t_end = vma->vm_end;
 442         return t_end;
 443 }
 444
 445 /*
 446  * Called with hugetlb fault mutex held.  Therefore, no more mappings to
 447  * this folio can be created while executing the routine.
 448  */
 449 static void hugetlb_unmap_file_folio(struct hstate *h,
 450                                         struct address_space *mapping,
 451                                         struct folio *folio, pgoff_t index)
 452 {
 453         struct rb_root_cached *root = &mapping->i_mmap;
 454         struct hugetlb_vma_lock *vma_lock;
 455         struct page *page = &folio->page;
 456         struct vm_area_struct *vma;
 457         unsigned long v_start;
 458         unsigned long v_end;
 459         pgoff_t start, end;
 460
 461         start = index * pages_per_huge_page(h);
 462         end = (index + 1) * pages_per_huge_page(h);
 463
 464         i_mmap_lock_write(mapping);
 465 retry:
 466         vma_lock = NULL;
 467         vma_interval_tree_foreach(vma, root, start, end - 1) {
 468                 v_start = vma_offset_start(vma, start);
 469                 v_end = vma_offset_end(vma, end);
 470
 471                 if (!hugetlb_vma_maps_page(vma, vma->vm_start + v_start, page))
 472                         continue;
 473
 474                 if (!hugetlb_vma_trylock_write(vma)) {
 475                         vma_lock = vma->vm_private_data;
 476                         /*
 477                          * If we can not get vma lock, we need to drop
 478                          * immap_sema and take locks in order.  First,
 479                          * take a ref on the vma_lock structure so that
 480                          * we can be guaranteed it will not go away when
 481                          * dropping immap_sema.
 482                          */
 483                         kref_get(&vma_lock->refs);
 484                         break;
 485                 }
 486
 487                 unmap_hugepage_range(vma, vma->vm_start + v_start, v_end,
 488                                 NULL, ZAP_FLAG_DROP_MARKER);
 489                 hugetlb_vma_unlock_write(vma);
 490         }
 491
 492         i_mmap_unlock_write(mapping);
 493
 494         if (vma_lock) {
 495                 /*
 496                  * Wait on vma_lock.  We know it is still valid as we have
 497                  * a reference.  We must 'open code' vma locking as we do
 498                  * not know if vma_lock is still attached to vma.
 499                  */
 500                 down_write(&vma_lock->rw_sema);
 501                 i_mmap_lock_write(mapping);
 502
 503                 vma = vma_lock->vma;
 504                 if (!vma) {
 505                         /*
 506                          * If lock is no longer attached to vma, then just
 507                          * unlock, drop our reference and retry looking for
 508                          * other vmas.
 509                          */
 510                         up_write(&vma_lock->rw_sema);
 511                         kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
 512                         goto retry;
 513                 }
 514
 515                 /*
 516                  * vma_lock is still attached to vma.  Check to see if vma
 517                  * still maps page and if so, unmap.
 518                  */
 519                 v_start = vma_offset_start(vma, start);
 520                 v_end = vma_offset_end(vma, end);
 521                 if (hugetlb_vma_maps_page(vma, vma->vm_start + v_start, page))
 522                         unmap_hugepage_range(vma, vma->vm_start + v_start,
 523                                                 v_end, NULL,
 524                                                 ZAP_FLAG_DROP_MARKER);
 525
 526                 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
 527                 hugetlb_vma_unlock_write(vma);
 528
 529                 goto retry;
 530         }
 531 }
 532
 533 static void
 534 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
 535                       zap_flags_t zap_flags)
 536 {
 537         struct vm_area_struct *vma;
 538
 539         /*
 540          * end == 0 indicates that the entire range after start should be
 541          * unmapped.  Note, end is exclusive, whereas the interval tree takes
 542          * an inclusive "last".
 543          */
 544         vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
 545                 unsigned long v_start;
 546                 unsigned long v_end;
 547
 548                 if (!hugetlb_vma_trylock_write(vma))
 549                         continue;
 550
 551                 v_start = vma_offset_start(vma, start);
 552                 v_end = vma_offset_end(vma, end);
 553
 554                 unmap_hugepage_range(vma, vma->vm_start + v_start, v_end,
 555                                      NULL, zap_flags);
 556
 557                 /*
 558                  * Note that vma lock only exists for shared/non-private
 559                  * vmas.  Therefore, lock is not held when calling
 560                  * unmap_hugepage_range for private vmas.
 561                  */
 562                 hugetlb_vma_unlock_write(vma);
 563         }
 564 }
 565
 566 /*
 567  * Called with hugetlb fault mutex held.
 568  * Returns true if page was actually removed, false otherwise.
 569  */
 570 static bool remove_inode_single_folio(struct hstate *h, struct inode *inode,
 571                                         struct address_space *mapping,
 572                                         struct folio *folio, pgoff_t index,
 573                                         bool truncate_op)
 574 {
 575         bool ret = false;
 576
 577         /*
 578          * If folio is mapped, it was faulted in after being
 579          * unmapped in caller.  Unmap (again) while holding
 580          * the fault mutex.  The mutex will prevent faults
 581          * until we finish removing the folio.
 582          */
 583         if (unlikely(folio_mapped(folio)))
 584                 hugetlb_unmap_file_folio(h, mapping, folio, index);
 585
 586         folio_lock(folio);
 587         /*
 588          * We must remove the folio from page cache before removing
 589          * the region/ reserve map (hugetlb_unreserve_pages).  In
 590          * rare out of memory conditions, removal of the region/reserve
 591          * map could fail.  Correspondingly, the subpool and global
 592          * reserve usage count can need to be adjusted.
 593          */
 594         VM_BUG_ON(HPageRestoreReserve(&folio->page));
 595         hugetlb_delete_from_page_cache(&folio->page);
 596         ret = true;
 597         if (!truncate_op) {
 598                 if (unlikely(hugetlb_unreserve_pages(inode, index,
 599                                                         index + 1, 1)))
 600                         hugetlb_fix_reserve_counts(inode);
 601         }
 602
 603         folio_unlock(folio);
 604         return ret;
 605 }
 606
 607 /*
 608  * remove_inode_hugepages handles two distinct cases: truncation and hole
 609  * punch.  There are subtle differences in operation for each case.
 610  *
 611  * truncation is indicated by end of range being LLONG_MAX
 612  *      In this case, we first scan the range and release found pages.
 613  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
 614  *      maps and global counts.  Page faults can race with truncation.
 615  *      During faults, hugetlb_no_page() checks i_size before page allocation,
 616  *      and again after obtaining page table lock.  It will 'back out'
 617  *      allocations in the truncated range.
 618  * hole punch is indicated if end is not LLONG_MAX
 619  *      In the hole punch case we scan the range and release found pages.
 620  *      Only when releasing a page is the associated region/reserve map
 621  *      deleted.  The region/reserve map for ranges without associated
 622  *      pages are not modified.  Page faults can race with hole punch.
 623  *      This is indicated if we find a mapped page.
 624  * Note: If the passed end of range value is beyond the end of file, but
 625  * not LLONG_MAX this routine still performs a hole punch operation.
 626  */
 627 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 628                                    loff_t lend)
 629 {
 630         struct hstate *h = hstate_inode(inode);
 631         struct address_space *mapping = &inode->i_data;
 632         const pgoff_t start = lstart >> huge_page_shift(h);
 633         const pgoff_t end = lend >> huge_page_shift(h);
 634         struct folio_batch fbatch;
 635         pgoff_t next, index;
 636         int i, freed = 0;
 637         bool truncate_op = (lend == LLONG_MAX);
 638
 639         folio_batch_init(&fbatch);
 640         next = start;
 641         while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
 642                 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
 643                         struct folio *folio = fbatch.folios[i];
 644                         u32 hash = 0;
 645
 646                         index = folio->index;
 647                         hash = hugetlb_fault_mutex_hash(mapping, index);
 648                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 649
 650                         /*
 651                          * Remove folio that was part of folio_batch.
 652                          */
 653                         if (remove_inode_single_folio(h, inode, mapping, folio,
 654                                                         index, truncate_op))
 655                                 freed++;
 656
 657                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 658                 }
 659                 folio_batch_release(&fbatch);
 660                 cond_resched();
 661         }
 662
 663         if (truncate_op)
 664                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 665 }
 666
 667 static void hugetlbfs_evict_inode(struct inode *inode)
 668 {
 669         struct resv_map *resv_map;
 670
 671         remove_inode_hugepages(inode, 0, LLONG_MAX);
 672
 673         /*
 674          * Get the resv_map from the address space embedded in the inode.
 675          * This is the address space which points to any resv_map allocated
 676          * at inode creation time.  If this is a device special inode,
 677          * i_mapping may not point to the original address space.
 678          */
 679         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
 680         /* Only regular and link inodes have associated reserve maps */
 681         if (resv_map)
 682                 resv_map_release(&resv_map->refs);
 683         clear_inode(inode);
 684 }
 685
 686 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 687 {
 688         pgoff_t pgoff;
 689         struct address_space *mapping = inode->i_mapping;
 690         struct hstate *h = hstate_inode(inode);
 691
 692         BUG_ON(offset & ~huge_page_mask(h));
 693         pgoff = offset >> PAGE_SHIFT;
 694
 695         i_size_write(inode, offset);
 696         i_mmap_lock_write(mapping);
 697         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 698                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
 699                                       ZAP_FLAG_DROP_MARKER);
 700         i_mmap_unlock_write(mapping);
 701         remove_inode_hugepages(inode, offset, LLONG_MAX);
 702 }
 703
 704 static void hugetlbfs_zero_partial_page(struct hstate *h,
 705                                         struct address_space *mapping,
 706                                         loff_t start,
 707                                         loff_t end)
 708 {
 709         pgoff_t idx = start >> huge_page_shift(h);
 710         struct folio *folio;
 711
 712         folio = filemap_lock_folio(mapping, idx);
 713         if (!folio)
 714                 return;
 715
 716         start = start & ~huge_page_mask(h);
 717         end = end & ~huge_page_mask(h);
 718         if (!end)
 719                 end = huge_page_size(h);
 720
 721         folio_zero_segment(folio, (size_t)start, (size_t)end);
 722
 723         folio_unlock(folio);
 724         folio_put(folio);
 725 }
 726
 727 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 728 {
 729         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 730         struct address_space *mapping = inode->i_mapping;
 731         struct hstate *h = hstate_inode(inode);
 732         loff_t hpage_size = huge_page_size(h);
 733         loff_t hole_start, hole_end;
 734
 735         /*
 736          * hole_start and hole_end indicate the full pages within the hole.
 737          */
 738         hole_start = round_up(offset, hpage_size);
 739         hole_end = round_down(offset + len, hpage_size);
 740
 741         inode_lock(inode);
 742
 743         /* protected by i_rwsem */
 744         if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 745                 inode_unlock(inode);
 746                 return -EPERM;
 747         }
 748
 749         i_mmap_lock_write(mapping);
 750
 751         /* If range starts before first full page, zero partial page. */
 752         if (offset < hole_start)
 753                 hugetlbfs_zero_partial_page(h, mapping,
 754                                 offset, min(offset + len, hole_start));
 755
 756         /* Unmap users of full pages in the hole. */
 757         if (hole_end > hole_start) {
 758                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 759                         hugetlb_vmdelete_list(&mapping->i_mmap,
 760                                               hole_start >> PAGE_SHIFT,
 761                                               hole_end >> PAGE_SHIFT, 0);
 762         }
 763
 764         /* If range extends beyond last full page, zero partial page. */
 765         if ((offset + len) > hole_end && (offset + len) > hole_start)
 766                 hugetlbfs_zero_partial_page(h, mapping,
 767                                 hole_end, offset + len);
 768
 769         i_mmap_unlock_write(mapping);
 770
 771         /* Remove full pages from the file. */
 772         if (hole_end > hole_start)
 773                 remove_inode_hugepages(inode, hole_start, hole_end);
 774
 775         inode_unlock(inode);
 776
 777         return 0;
 778 }
 779
 780 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 781                                 loff_t len)
 782 {
 783         struct inode *inode = file_inode(file);
 784         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 785         struct address_space *mapping = inode->i_mapping;
 786         struct hstate *h = hstate_inode(inode);
 787         struct vm_area_struct pseudo_vma;
 788         struct mm_struct *mm = current->mm;
 789         loff_t hpage_size = huge_page_size(h);
 790         unsigned long hpage_shift = huge_page_shift(h);
 791         pgoff_t start, index, end;
 792         int error;
 793         u32 hash;
 794
 795         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 796                 return -EOPNOTSUPP;
 797
 798         if (mode & FALLOC_FL_PUNCH_HOLE)
 799                 return hugetlbfs_punch_hole(inode, offset, len);
 800
 801         /*
 802          * Default preallocate case.
 803          * For this range, start is rounded down and end is rounded up
 804          * as well as being converted to page offsets.
 805          */
 806         start = offset >> hpage_shift;
 807         end = (offset + len + hpage_size - 1) >> hpage_shift;
 808
 809         inode_lock(inode);
 810
 811         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 812         error = inode_newsize_ok(inode, offset + len);
 813         if (error)
 814                 goto out;
 815
 816         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 817                 error = -EPERM;
 818                 goto out;
 819         }
 820
 821         /*
 822          * Initialize a pseudo vma as this is required by the huge page
 823          * allocation routines.  If NUMA is configured, use page index
 824          * as input to create an allocation policy.
 825          */
 826         vma_init(&pseudo_vma, mm);
 827         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 828         pseudo_vma.vm_file = file;
 829
 830         for (index = start; index < end; index++) {
 831                 /*
 832                  * This is supposed to be the vaddr where the page is being
 833                  * faulted in, but we have no vaddr here.
 834                  */
 835                 struct page *page;
 836                 unsigned long addr;
 837
 838                 cond_resched();
 839
 840                 /*
 841                  * fallocate(2) manpage permits EINTR; we may have been
 842                  * interrupted because we are using up too much memory.
 843                  */
 844                 if (signal_pending(current)) {
 845                         error = -EINTR;
 846                         break;
 847                 }
 848
 849                 /* Set numa allocation policy based on index */
 850                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 851
 852                 /* addr is the offset within the file (zero based) */
 853                 addr = index * hpage_size;
 854
 855                 /* mutex taken here, fault path and hole punch */
 856                 hash = hugetlb_fault_mutex_hash(mapping, index);
 857                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 858
 859                 /* See if already present in mapping to avoid alloc/free */
 860                 page = find_get_page(mapping, index);
 861                 if (page) {
 862                         put_page(page);
 863                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 864                         hugetlb_drop_vma_policy(&pseudo_vma);
 865                         continue;
 866                 }
 867
 868                 /*
 869                  * Allocate page without setting the avoid_reserve argument.
 870                  * There certainly are no reserves associated with the
 871                  * pseudo_vma.  However, there could be shared mappings with
 872                  * reserves for the file at the inode level.  If we fallocate
 873                  * pages in these areas, we need to consume the reserves
 874                  * to keep reservation accounting consistent.
 875                  */
 876                 page = alloc_huge_page(&pseudo_vma, addr, 0);
 877                 hugetlb_drop_vma_policy(&pseudo_vma);
 878                 if (IS_ERR(page)) {
 879                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 880                         error = PTR_ERR(page);
 881                         goto out;
 882                 }
 883                 clear_huge_page(page, addr, pages_per_huge_page(h));
 884                 __SetPageUptodate(page);
 885                 error = hugetlb_add_to_page_cache(page, mapping, index);
 886                 if (unlikely(error)) {
 887                         restore_reserve_on_error(h, &pseudo_vma, addr, page);
 888                         put_page(page);
 889                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 890                         goto out;
 891                 }
 892
 893                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 894
 895                 SetHPageMigratable(page);
 896                 /*
 897                  * unlock_page because locked by hugetlb_add_to_page_cache()
 898                  * put_page() due to reference from alloc_huge_page()
 899                  */
 900                 unlock_page(page);
 901                 put_page(page);
 902         }
 903
 904         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 905                 i_size_write(inode, offset + len);
 906         inode->i_ctime = current_time(inode);
 907 out:
 908         inode_unlock(inode);
 909         return error;
 910 }
 911
 912 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
 913                              struct dentry *dentry, struct iattr *attr)
 914 {
 915         struct inode *inode = d_inode(dentry);
 916         struct hstate *h = hstate_inode(inode);
 917         int error;
 918         unsigned int ia_valid = attr->ia_valid;
 919         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 920
 921         error = setattr_prepare(&init_user_ns, dentry, attr);
 922         if (error)
 923                 return error;
 924
 925         if (ia_valid & ATTR_SIZE) {
 926                 loff_t oldsize = inode->i_size;
 927                 loff_t newsize = attr->ia_size;
 928
 929                 if (newsize & ~huge_page_mask(h))
 930                         return -EINVAL;
 931                 /* protected by i_rwsem */
 932                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 933                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 934                         return -EPERM;
 935                 hugetlb_vmtruncate(inode, newsize);
 936         }
 937
 938         setattr_copy(&init_user_ns, inode, attr);
 939         mark_inode_dirty(inode);
 940         return 0;
 941 }
 942
 943 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 944                                         struct hugetlbfs_fs_context *ctx)
 945 {
 946         struct inode *inode;
 947
 948         inode = new_inode(sb);
 949         if (inode) {
 950                 inode->i_ino = get_next_ino();
 951                 inode->i_mode = S_IFDIR | ctx->mode;
 952                 inode->i_uid = ctx->uid;
 953                 inode->i_gid = ctx->gid;
 954                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 955                 inode->i_op = &hugetlbfs_dir_inode_operations;
 956                 inode->i_fop = &simple_dir_operations;
 957                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 958                 inc_nlink(inode);
 959                 lockdep_annotate_inode_mutex_key(inode);
 960         }
 961         return inode;
 962 }
 963
 964 /*
 965  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 966  * be taken from reclaim -- unlike regular filesystems. This needs an
 967  * annotation because huge_pmd_share() does an allocation under hugetlb's
 968  * i_mmap_rwsem.
 969  */
 970 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 971
 972 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 973                                         struct inode *dir,
 974                                         umode_t mode, dev_t dev)
 975 {
 976         struct inode *inode;
 977         struct resv_map *resv_map = NULL;
 978
 979         /*
 980          * Reserve maps are only needed for inodes that can have associated
 981          * page allocations.
 982          */
 983         if (S_ISREG(mode) || S_ISLNK(mode)) {
 984                 resv_map = resv_map_alloc();
 985                 if (!resv_map)
 986                         return NULL;
 987         }
 988
 989         inode = new_inode(sb);
 990         if (inode) {
 991                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 992
 993                 inode->i_ino = get_next_ino();
 994                 inode_init_owner(&init_user_ns, inode, dir, mode);
 995                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 996                                 &hugetlbfs_i_mmap_rwsem_key);
 997                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 998                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 999                 inode->i_mapping->private_data = resv_map;
1000                 info->seals = F_SEAL_SEAL;
1001                 switch (mode & S_IFMT) {
1002                 default:
1003                         init_special_inode(inode, mode, dev);
1004                         break;
1005                 case S_IFREG:
1006                         inode->i_op = &hugetlbfs_inode_operations;
1007                         inode->i_fop = &hugetlbfs_file_operations;
1008                         break;
1009                 case S_IFDIR:
1010                         inode->i_op = &hugetlbfs_dir_inode_operations;
1011                         inode->i_fop = &simple_dir_operations;
1012
1013                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
1014                         inc_nlink(inode);
1015                         break;
1016                 case S_IFLNK:
1017                         inode->i_op = &page_symlink_inode_operations;
1018                         inode_nohighmem(inode);
1019                         break;
1020                 }
1021                 lockdep_annotate_inode_mutex_key(inode);
1022         } else {
1023                 if (resv_map)
1024                         kref_put(&resv_map->refs, resv_map_release);
1025         }
1026
1027         return inode;
1028 }
1029
1030 /*
1031  * File creation. Allocate an inode, and we're done..
1032  */
1033 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
1034                            struct dentry *dentry, umode_t mode, dev_t dev)
1035 {
1036         struct inode *inode;
1037
1038         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
1039         if (!inode)
1040                 return -ENOSPC;
1041         dir->i_ctime = dir->i_mtime = current_time(dir);
1042         d_instantiate(dentry, inode);
1043         dget(dentry);/* Extra count - pin the dentry in core */
1044         return 0;
1045 }
1046
1047 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
1048                            struct dentry *dentry, umode_t mode)
1049 {
1050         int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
1051                                      mode | S_IFDIR, 0);
1052         if (!retval)
1053                 inc_nlink(dir);
1054         return retval;
1055 }
1056
1057 static int hugetlbfs_create(struct user_namespace *mnt_userns,
1058                             struct inode *dir, struct dentry *dentry,
1059                             umode_t mode, bool excl)
1060 {
1061         return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
1062 }
1063
1064 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
1065                              struct inode *dir, struct file *file,
1066                              umode_t mode)
1067 {
1068         struct inode *inode;
1069
1070         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode | S_IFREG, 0);
1071         if (!inode)
1072                 return -ENOSPC;
1073         dir->i_ctime = dir->i_mtime = current_time(dir);
1074         d_tmpfile(file, inode);
1075         return finish_open_simple(file, 0);
1076 }
1077
1078 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
1079                              struct inode *dir, struct dentry *dentry,
1080                              const char *symname)
1081 {
1082         struct inode *inode;
1083         int error = -ENOSPC;
1084
1085         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
1086         if (inode) {
1087                 int l = strlen(symname)+1;
1088                 error = page_symlink(inode, symname, l);
1089                 if (!error) {
1090                         d_instantiate(dentry, inode);
1091                         dget(dentry);
1092                 } else
1093                         iput(inode);
1094         }
1095         dir->i_ctime = dir->i_mtime = current_time(dir);
1096
1097         return error;
1098 }
1099
1100 #ifdef CONFIG_MIGRATION
1101 static int hugetlbfs_migrate_folio(struct address_space *mapping,
1102                                 struct folio *dst, struct folio *src,
1103                                 enum migrate_mode mode)
1104 {
1105         int rc;
1106
1107         rc = migrate_huge_page_move_mapping(mapping, dst, src);
1108         if (rc != MIGRATEPAGE_SUCCESS)
1109                 return rc;
1110
1111         if (hugetlb_page_subpool(&src->page)) {
1112                 hugetlb_set_page_subpool(&dst->page,
1113                                         hugetlb_page_subpool(&src->page));
1114                 hugetlb_set_page_subpool(&src->page, NULL);
1115         }
1116
1117         if (mode != MIGRATE_SYNC_NO_COPY)
1118                 folio_migrate_copy(dst, src);
1119         else
1120                 folio_migrate_flags(dst, src);
1121
1122         return MIGRATEPAGE_SUCCESS;
1123 }
1124 #else
1125 #define hugetlbfs_migrate_folio NULL
1126 #endif
1127
1128 static int hugetlbfs_error_remove_page(struct address_space *mapping,
1129                                 struct page *page)
1130 {
1131         return 0;
1132 }
1133
1134 /*
1135  * Display the mount options in /proc/mounts.
1136  */
1137 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1138 {
1139         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1140         struct hugepage_subpool *spool = sbinfo->spool;
1141         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1142         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1143         char mod;
1144
1145         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1146                 seq_printf(m, ",uid=%u",
1147                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1148         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1149                 seq_printf(m, ",gid=%u",
1150                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1151         if (sbinfo->mode != 0755)
1152                 seq_printf(m, ",mode=%o", sbinfo->mode);
1153         if (sbinfo->max_inodes != -1)
1154                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1155
1156         hpage_size /= 1024;
1157         mod = 'K';
1158         if (hpage_size >= 1024) {
1159                 hpage_size /= 1024;
1160                 mod = 'M';
1161         }
1162         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1163         if (spool) {
1164                 if (spool->max_hpages != -1)
1165                         seq_printf(m, ",size=%llu",
1166                                    (unsigned long long)spool->max_hpages << hpage_shift);
1167                 if (spool->min_hpages != -1)
1168                         seq_printf(m, ",min_size=%llu",
1169                                    (unsigned long long)spool->min_hpages << hpage_shift);
1170         }
1171         return 0;
1172 }
1173
1174 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1175 {
1176         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1177         struct hstate *h = hstate_inode(d_inode(dentry));
1178
1179         buf->f_type = HUGETLBFS_MAGIC;
1180         buf->f_bsize = huge_page_size(h);
1181         if (sbinfo) {
1182                 spin_lock(&sbinfo->stat_lock);
1183                 /* If no limits set, just report 0 or -1 for max/free/used
1184                  * blocks, like simple_statfs() */
1185                 if (sbinfo->spool) {
1186                         long free_pages;
1187
1188                         spin_lock_irq(&sbinfo->spool->lock);
1189                         buf->f_blocks = sbinfo->spool->max_hpages;
1190                         free_pages = sbinfo->spool->max_hpages
1191                                 - sbinfo->spool->used_hpages;
1192                         buf->f_bavail = buf->f_bfree = free_pages;
1193                         spin_unlock_irq(&sbinfo->spool->lock);
1194                         buf->f_files = sbinfo->max_inodes;
1195                         buf->f_ffree = sbinfo->free_inodes;
1196                 }
1197                 spin_unlock(&sbinfo->stat_lock);
1198         }
1199         buf->f_namelen = NAME_MAX;
1200         return 0;
1201 }
1202
1203 static void hugetlbfs_put_super(struct super_block *sb)
1204 {
1205         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1206
1207         if (sbi) {
1208                 sb->s_fs_info = NULL;
1209
1210                 if (sbi->spool)
1211                         hugepage_put_subpool(sbi->spool);
1212
1213                 kfree(sbi);
1214         }
1215 }
1216
1217 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1218 {
1219         if (sbinfo->free_inodes >= 0) {
1220                 spin_lock(&sbinfo->stat_lock);
1221                 if (unlikely(!sbinfo->free_inodes)) {
1222                         spin_unlock(&sbinfo->stat_lock);
1223                         return 0;
1224                 }
1225                 sbinfo->free_inodes--;
1226                 spin_unlock(&sbinfo->stat_lock);
1227         }
1228
1229         return 1;
1230 }
1231
1232 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1233 {
1234         if (sbinfo->free_inodes >= 0) {
1235                 spin_lock(&sbinfo->stat_lock);
1236                 sbinfo->free_inodes++;
1237                 spin_unlock(&sbinfo->stat_lock);
1238         }
1239 }
1240
1241
1242 static struct kmem_cache *hugetlbfs_inode_cachep;
1243
1244 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1245 {
1246         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1247         struct hugetlbfs_inode_info *p;
1248
1249         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1250                 return NULL;
1251         p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1252         if (unlikely(!p)) {
1253                 hugetlbfs_inc_free_inodes(sbinfo);
1254                 return NULL;
1255         }
1256
1257         /*
1258          * Any time after allocation, hugetlbfs_destroy_inode can be called
1259          * for the inode.  mpol_free_shared_policy is unconditionally called
1260          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1261          * in case of a quick call to destroy.
1262          *
1263          * Note that the policy is initialized even if we are creating a
1264          * private inode.  This simplifies hugetlbfs_destroy_inode.
1265          */
1266         mpol_shared_policy_init(&p->policy, NULL);
1267
1268         return &p->vfs_inode;
1269 }
1270
1271 static void hugetlbfs_free_inode(struct inode *inode)
1272 {
1273         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1274 }
1275
1276 static void hugetlbfs_destroy_inode(struct inode *inode)
1277 {
1278         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1279         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1280 }
1281
1282 static const struct address_space_operations hugetlbfs_aops = {
1283         .write_begin    = hugetlbfs_write_begin,
1284         .write_end      = hugetlbfs_write_end,
1285         .dirty_folio    = noop_dirty_folio,
1286         .migrate_folio  = hugetlbfs_migrate_folio,
1287         .error_remove_page      = hugetlbfs_error_remove_page,
1288 };
1289
1290
1291 static void init_once(void *foo)
1292 {
1293         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1294
1295         inode_init_once(&ei->vfs_inode);
1296 }
1297
1298 const struct file_operations hugetlbfs_file_operations = {
1299         .read_iter              = hugetlbfs_read_iter,
1300         .mmap                   = hugetlbfs_file_mmap,
1301         .fsync                  = noop_fsync,
1302         .get_unmapped_area      = hugetlb_get_unmapped_area,
1303         .llseek                 = default_llseek,
1304         .fallocate              = hugetlbfs_fallocate,
1305 };
1306
1307 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1308         .create         = hugetlbfs_create,
1309         .lookup         = simple_lookup,
1310         .link           = simple_link,
1311         .unlink         = simple_unlink,
1312         .symlink        = hugetlbfs_symlink,
1313         .mkdir          = hugetlbfs_mkdir,
1314         .rmdir          = simple_rmdir,
1315         .mknod          = hugetlbfs_mknod,
1316         .rename         = simple_rename,
1317         .setattr        = hugetlbfs_setattr,
1318         .tmpfile        = hugetlbfs_tmpfile,
1319 };
1320
1321 static const struct inode_operations hugetlbfs_inode_operations = {
1322         .setattr        = hugetlbfs_setattr,
1323 };
1324
1325 static const struct super_operations hugetlbfs_ops = {
1326         .alloc_inode    = hugetlbfs_alloc_inode,
1327         .free_inode     = hugetlbfs_free_inode,
1328         .destroy_inode  = hugetlbfs_destroy_inode,
1329         .evict_inode    = hugetlbfs_evict_inode,
1330         .statfs         = hugetlbfs_statfs,
1331         .put_super      = hugetlbfs_put_super,
1332         .show_options   = hugetlbfs_show_options,
1333 };
1334
1335 /*
1336  * Convert size option passed from command line to number of huge pages
1337  * in the pool specified by hstate.  Size option could be in bytes
1338  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1339  */
1340 static long
1341 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1342                          enum hugetlbfs_size_type val_type)
1343 {
1344         if (val_type == NO_SIZE)
1345                 return -1;
1346
1347         if (val_type == SIZE_PERCENT) {
1348                 size_opt <<= huge_page_shift(h);
1349                 size_opt *= h->max_huge_pages;
1350                 do_div(size_opt, 100);
1351         }
1352
1353         size_opt >>= huge_page_shift(h);
1354         return size_opt;
1355 }
1356
1357 /*
1358  * Parse one mount parameter.
1359  */
1360 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1361 {
1362         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1363         struct fs_parse_result result;
1364         struct hstate *h;
1365         char *rest;
1366         unsigned long ps;
1367         int opt;
1368
1369         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1370         if (opt < 0)
1371                 return opt;
1372
1373         switch (opt) {
1374         case Opt_uid:
1375                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1376                 if (!uid_valid(ctx->uid))
1377                         goto bad_val;
1378                 return 0;
1379
1380         case Opt_gid:
1381                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1382                 if (!gid_valid(ctx->gid))
1383                         goto bad_val;
1384                 return 0;
1385
1386         case Opt_mode:
1387                 ctx->mode = result.uint_32 & 01777U;
1388                 return 0;
1389
1390         case Opt_size:
1391                 /* memparse() will accept a K/M/G without a digit */
1392                 if (!param->string || !isdigit(param->string[0]))
1393                         goto bad_val;
1394                 ctx->max_size_opt = memparse(param->string, &rest);
1395                 ctx->max_val_type = SIZE_STD;
1396                 if (*rest == '%')
1397                         ctx->max_val_type = SIZE_PERCENT;
1398                 return 0;
1399
1400         case Opt_nr_inodes:
1401                 /* memparse() will accept a K/M/G without a digit */
1402                 if (!param->string || !isdigit(param->string[0]))
1403                         goto bad_val;
1404                 ctx->nr_inodes = memparse(param->string, &rest);
1405                 return 0;
1406
1407         case Opt_pagesize:
1408                 ps = memparse(param->string, &rest);
1409                 h = size_to_hstate(ps);
1410                 if (!h) {
1411                         pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1412                         return -EINVAL;
1413                 }
1414                 ctx->hstate = h;
1415                 return 0;
1416
1417         case Opt_min_size:
1418                 /* memparse() will accept a K/M/G without a digit */
1419                 if (!param->string || !isdigit(param->string[0]))
1420                         goto bad_val;
1421                 ctx->min_size_opt = memparse(param->string, &rest);
1422                 ctx->min_val_type = SIZE_STD;
1423                 if (*rest == '%')
1424                         ctx->min_val_type = SIZE_PERCENT;
1425                 return 0;
1426
1427         default:
1428                 return -EINVAL;
1429         }
1430
1431 bad_val:
1432         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1433                       param->string, param->key);
1434 }
1435
1436 /*
1437  * Validate the parsed options.
1438  */
1439 static int hugetlbfs_validate(struct fs_context *fc)
1440 {
1441         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1442
1443         /*
1444          * Use huge page pool size (in hstate) to convert the size
1445          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1446          */
1447         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1448                                                    ctx->max_size_opt,
1449                                                    ctx->max_val_type);
1450         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1451                                                    ctx->min_size_opt,
1452                                                    ctx->min_val_type);
1453
1454         /*
1455          * If max_size was specified, then min_size must be smaller
1456          */
1457         if (ctx->max_val_type > NO_SIZE &&
1458             ctx->min_hpages > ctx->max_hpages) {
1459                 pr_err("Minimum size can not be greater than maximum size\n");
1460                 return -EINVAL;
1461         }
1462
1463         return 0;
1464 }
1465
1466 static int
1467 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1468 {
1469         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1470         struct hugetlbfs_sb_info *sbinfo;
1471
1472         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1473         if (!sbinfo)
1474                 return -ENOMEM;
1475         sb->s_fs_info = sbinfo;
1476         spin_lock_init(&sbinfo->stat_lock);
1477         sbinfo->hstate          = ctx->hstate;
1478         sbinfo->max_inodes      = ctx->nr_inodes;
1479         sbinfo->free_inodes     = ctx->nr_inodes;
1480         sbinfo->spool           = NULL;
1481         sbinfo->uid             = ctx->uid;
1482         sbinfo->gid             = ctx->gid;
1483         sbinfo->mode            = ctx->mode;
1484
1485         /*
1486          * Allocate and initialize subpool if maximum or minimum size is
1487          * specified.  Any needed reservations (for minimum size) are taken
1488          * when the subpool is created.
1489          */
1490         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1491                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1492                                                      ctx->max_hpages,
1493                                                      ctx->min_hpages);
1494                 if (!sbinfo->spool)
1495                         goto out_free;
1496         }
1497         sb->s_maxbytes = MAX_LFS_FILESIZE;
1498         sb->s_blocksize = huge_page_size(ctx->hstate);
1499         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1500         sb->s_magic = HUGETLBFS_MAGIC;
1501         sb->s_op = &hugetlbfs_ops;
1502         sb->s_time_gran = 1;
1503
1504         /*
1505          * Due to the special and limited functionality of hugetlbfs, it does
1506          * not work well as a stacking filesystem.
1507          */
1508         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1509         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1510         if (!sb->s_root)
1511                 goto out_free;
1512         return 0;
1513 out_free:
1514         kfree(sbinfo->spool);
1515         kfree(sbinfo);
1516         return -ENOMEM;
1517 }
1518
1519 static int hugetlbfs_get_tree(struct fs_context *fc)
1520 {
1521         int err = hugetlbfs_validate(fc);
1522         if (err)
1523                 return err;
1524         return get_tree_nodev(fc, hugetlbfs_fill_super);
1525 }
1526
1527 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1528 {
1529         kfree(fc->fs_private);
1530 }
1531
1532 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1533         .free           = hugetlbfs_fs_context_free,
1534         .parse_param    = hugetlbfs_parse_param,
1535         .get_tree       = hugetlbfs_get_tree,
1536 };
1537
1538 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1539 {
1540         struct hugetlbfs_fs_context *ctx;
1541
1542         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1543         if (!ctx)
1544                 return -ENOMEM;
1545
1546         ctx->max_hpages = -1; /* No limit on size by default */
1547         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1548         ctx->uid        = current_fsuid();
1549         ctx->gid        = current_fsgid();
1550         ctx->mode       = 0755;
1551         ctx->hstate     = &default_hstate;
1552         ctx->min_hpages = -1; /* No default minimum size */
1553         ctx->max_val_type = NO_SIZE;
1554         ctx->min_val_type = NO_SIZE;
1555         fc->fs_private = ctx;
1556         fc->ops = &hugetlbfs_fs_context_ops;
1557         return 0;
1558 }
1559
1560 static struct file_system_type hugetlbfs_fs_type = {
1561         .name                   = "hugetlbfs",
1562         .init_fs_context        = hugetlbfs_init_fs_context,
1563         .parameters             = hugetlb_fs_parameters,
1564         .kill_sb                = kill_litter_super,
1565 };
1566
1567 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1568
1569 static int can_do_hugetlb_shm(void)
1570 {
1571         kgid_t shm_group;
1572         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1573         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1574 }
1575
1576 static int get_hstate_idx(int page_size_log)
1577 {
1578         struct hstate *h = hstate_sizelog(page_size_log);
1579
1580         if (!h)
1581                 return -1;
1582         return hstate_index(h);
1583 }
1584
1585 /*
1586  * Note that size should be aligned to proper hugepage size in caller side,
1587  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1588  */
1589 struct file *hugetlb_file_setup(const char *name, size_t size,
1590                                 vm_flags_t acctflag, int creat_flags,
1591                                 int page_size_log)
1592 {
1593         struct inode *inode;
1594         struct vfsmount *mnt;
1595         int hstate_idx;
1596         struct file *file;
1597
1598         hstate_idx = get_hstate_idx(page_size_log);
1599         if (hstate_idx < 0)
1600                 return ERR_PTR(-ENODEV);
1601
1602         mnt = hugetlbfs_vfsmount[hstate_idx];
1603         if (!mnt)
1604                 return ERR_PTR(-ENOENT);
1605
1606         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1607                 struct ucounts *ucounts = current_ucounts();
1608
1609                 if (user_shm_lock(size, ucounts)) {
1610                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1611                                 current->comm, current->pid);
1612                         user_shm_unlock(size, ucounts);
1613                 }
1614                 return ERR_PTR(-EPERM);
1615         }
1616
1617         file = ERR_PTR(-ENOSPC);
1618         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1619         if (!inode)
1620                 goto out;
1621         if (creat_flags == HUGETLB_SHMFS_INODE)
1622                 inode->i_flags |= S_PRIVATE;
1623
1624         inode->i_size = size;
1625         clear_nlink(inode);
1626
1627         if (!hugetlb_reserve_pages(inode, 0,
1628                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1629                         acctflag))
1630                 file = ERR_PTR(-ENOMEM);
1631         else
1632                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1633                                         &hugetlbfs_file_operations);
1634         if (!IS_ERR(file))
1635                 return file;
1636
1637         iput(inode);
1638 out:
1639         return file;
1640 }
1641
1642 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1643 {
1644         struct fs_context *fc;
1645         struct vfsmount *mnt;
1646
1647         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1648         if (IS_ERR(fc)) {
1649                 mnt = ERR_CAST(fc);
1650         } else {
1651                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1652                 ctx->hstate = h;
1653                 mnt = fc_mount(fc);
1654                 put_fs_context(fc);
1655         }
1656         if (IS_ERR(mnt))
1657                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1658                        huge_page_size(h) / SZ_1K);
1659         return mnt;
1660 }
1661
1662 static int __init init_hugetlbfs_fs(void)
1663 {
1664         struct vfsmount *mnt;
1665         struct hstate *h;
1666         int error;
1667         int i;
1668
1669         if (!hugepages_supported()) {
1670                 pr_info("disabling because there are no supported hugepage sizes\n");
1671                 return -ENOTSUPP;
1672         }
1673
1674         error = -ENOMEM;
1675         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1676                                         sizeof(struct hugetlbfs_inode_info),
1677                                         0, SLAB_ACCOUNT, init_once);
1678         if (hugetlbfs_inode_cachep == NULL)
1679                 goto out;
1680
1681         error = register_filesystem(&hugetlbfs_fs_type);
1682         if (error)
1683                 goto out_free;
1684
1685         /* default hstate mount is required */
1686         mnt = mount_one_hugetlbfs(&default_hstate);
1687         if (IS_ERR(mnt)) {
1688                 error = PTR_ERR(mnt);
1689                 goto out_unreg;
1690         }
1691         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1692
1693         /* other hstates are optional */
1694         i = 0;
1695         for_each_hstate(h) {
1696                 if (i == default_hstate_idx) {
1697                         i++;
1698                         continue;
1699                 }
1700
1701                 mnt = mount_one_hugetlbfs(h);
1702                 if (IS_ERR(mnt))
1703                         hugetlbfs_vfsmount[i] = NULL;
1704                 else
1705                         hugetlbfs_vfsmount[i] = mnt;
1706                 i++;
1707         }
1708
1709         return 0;
1710
1711  out_unreg:
1712         (void)unregister_filesystem(&hugetlbfs_fs_type);
1713  out_free:
1714         kmem_cache_destroy(hugetlbfs_inode_cachep);
1715  out:
1716         return error;
1717 }
1718 fs_initcall(init_hugetlbfs_fs)