fs/hugetlbfs/inode.c

   1 /*
   2  * hugetlbpage-backed filesystem.  Based on ramfs.
   3  *
   4  * Nadia Yvette Chambers, 2002
   5  *
   6  * Copyright (C) 2002 Linus Torvalds.
   7  * License: GPL
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/thread_info.h>
  13 #include <asm/current.h>
  14 #include <linux/falloc.h>
  15 #include <linux/fs.h>
  16 #include <linux/mount.h>
  17 #include <linux/file.h>
  18 #include <linux/kernel.h>
  19 #include <linux/writeback.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/highmem.h>
  22 #include <linux/init.h>
  23 #include <linux/string.h>
  24 #include <linux/capability.h>
  25 #include <linux/ctype.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/hugetlb.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/fs_parser.h>
  30 #include <linux/mman.h>
  31 #include <linux/slab.h>
  32 #include <linux/dnotify.h>
  33 #include <linux/statfs.h>
  34 #include <linux/security.h>
  35 #include <linux/magic.h>
  36 #include <linux/migrate.h>
  37 #include <linux/uio.h>
  38
  39 #include <linux/uaccess.h>
  40 #include <linux/sched/mm.h>
  41
  42 static const struct address_space_operations hugetlbfs_aops;
  43 const struct file_operations hugetlbfs_file_operations;
  44 static const struct inode_operations hugetlbfs_dir_inode_operations;
  45 static const struct inode_operations hugetlbfs_inode_operations;
  46
  47 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
  48
  49 struct hugetlbfs_fs_context {
  50         struct hstate           *hstate;
  51         unsigned long long      max_size_opt;
  52         unsigned long long      min_size_opt;
  53         long                    max_hpages;
  54         long                    nr_inodes;
  55         long                    min_hpages;
  56         enum hugetlbfs_size_type max_val_type;
  57         enum hugetlbfs_size_type min_val_type;
  58         kuid_t                  uid;
  59         kgid_t                  gid;
  60         umode_t                 mode;
  61 };
  62
  63 int sysctl_hugetlb_shm_group;
  64
  65 enum hugetlb_param {
  66         Opt_gid,
  67         Opt_min_size,
  68         Opt_mode,
  69         Opt_nr_inodes,
  70         Opt_pagesize,
  71         Opt_size,
  72         Opt_uid,
  73 };
  74
  75 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
  76         fsparam_u32   ("gid",           Opt_gid),
  77         fsparam_string("min_size",      Opt_min_size),
  78         fsparam_u32oct("mode",          Opt_mode),
  79         fsparam_string("nr_inodes",     Opt_nr_inodes),
  80         fsparam_string("pagesize",      Opt_pagesize),
  81         fsparam_string("size",          Opt_size),
  82         fsparam_u32   ("uid",           Opt_uid),
  83         {}
  84 };
  85
  86 #ifdef CONFIG_NUMA
  87 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  88                                         struct inode *inode, pgoff_t index)
  89 {
  90         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  91                                                         index);
  92 }
  93
  94 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  95 {
  96         mpol_cond_put(vma->vm_policy);
  97 }
  98 #else
  99 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
 100                                         struct inode *inode, pgoff_t index)
 101 {
 102 }
 103
 104 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
 105 {
 106 }
 107 #endif
 108
 109 /*
 110  * Mask used when checking the page offset value passed in via system
 111  * calls.  This value will be converted to a loff_t which is signed.
 112  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 113  * value.  The extra bit (- 1 in the shift value) is to take the sign
 114  * bit into account.
 115  */
 116 #define PGOFF_LOFFT_MAX \
 117         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 118
 119 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 120 {
 121         struct inode *inode = file_inode(file);
 122         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 123         loff_t len, vma_len;
 124         int ret;
 125         struct hstate *h = hstate_file(file);
 126
 127         /*
 128          * vma address alignment (but not the pgoff alignment) has
 129          * already been checked by prepare_hugepage_range.  If you add
 130          * any error returns here, do so after setting VM_HUGETLB, so
 131          * is_vm_hugetlb_page tests below unmap_region go the right
 132          * way when do_mmap unwinds (may be important on powerpc
 133          * and ia64).
 134          */
 135         vm_flags_set(vma, VM_HUGETLB | VM_DONTEXPAND);
 136         vma->vm_ops = &hugetlb_vm_ops;
 137
 138         ret = seal_check_future_write(info->seals, vma);
 139         if (ret)
 140                 return ret;
 141
 142         /*
 143          * page based offset in vm_pgoff could be sufficiently large to
 144          * overflow a loff_t when converted to byte offset.  This can
 145          * only happen on architectures where sizeof(loff_t) ==
 146          * sizeof(unsigned long).  So, only check in those instances.
 147          */
 148         if (sizeof(unsigned long) == sizeof(loff_t)) {
 149                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 150                         return -EINVAL;
 151         }
 152
 153         /* must be huge page aligned */
 154         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 155                 return -EINVAL;
 156
 157         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 158         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 159         /* check for overflow */
 160         if (len < vma_len)
 161                 return -EINVAL;
 162
 163         inode_lock(inode);
 164         file_accessed(file);
 165
 166         ret = -ENOMEM;
 167         if (!hugetlb_reserve_pages(inode,
 168                                 vma->vm_pgoff >> huge_page_order(h),
 169                                 len >> huge_page_shift(h), vma,
 170                                 vma->vm_flags))
 171                 goto out;
 172
 173         ret = 0;
 174         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 175                 i_size_write(inode, len);
 176 out:
 177         inode_unlock(inode);
 178
 179         return ret;
 180 }
 181
 182 /*
 183  * Called under mmap_write_lock(mm).
 184  */
 185
 186 static unsigned long
 187 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
 188                 unsigned long len, unsigned long pgoff, unsigned long flags)
 189 {
 190         struct hstate *h = hstate_file(file);
 191         struct vm_unmapped_area_info info;
 192
 193         info.flags = 0;
 194         info.length = len;
 195         info.low_limit = current->mm->mmap_base;
 196         info.high_limit = arch_get_mmap_end(addr, len, flags);
 197         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 198         info.align_offset = 0;
 199         return vm_unmapped_area(&info);
 200 }
 201
 202 static unsigned long
 203 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
 204                 unsigned long len, unsigned long pgoff, unsigned long flags)
 205 {
 206         struct hstate *h = hstate_file(file);
 207         struct vm_unmapped_area_info info;
 208
 209         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 210         info.length = len;
 211         info.low_limit = PAGE_SIZE;
 212         info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
 213         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 214         info.align_offset = 0;
 215         addr = vm_unmapped_area(&info);
 216
 217         /*
 218          * A failed mmap() very likely causes application failure,
 219          * so fall back to the bottom-up function here. This scenario
 220          * can happen with large stack limits and large mmap()
 221          * allocations.
 222          */
 223         if (unlikely(offset_in_page(addr))) {
 224                 VM_BUG_ON(addr != -ENOMEM);
 225                 info.flags = 0;
 226                 info.low_limit = current->mm->mmap_base;
 227                 info.high_limit = arch_get_mmap_end(addr, len, flags);
 228                 addr = vm_unmapped_area(&info);
 229         }
 230
 231         return addr;
 232 }
 233
 234 unsigned long
 235 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 236                                   unsigned long len, unsigned long pgoff,
 237                                   unsigned long flags)
 238 {
 239         struct mm_struct *mm = current->mm;
 240         struct vm_area_struct *vma;
 241         struct hstate *h = hstate_file(file);
 242         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
 243
 244         if (len & ~huge_page_mask(h))
 245                 return -EINVAL;
 246         if (len > TASK_SIZE)
 247                 return -ENOMEM;
 248
 249         if (flags & MAP_FIXED) {
 250                 if (prepare_hugepage_range(file, addr, len))
 251                         return -EINVAL;
 252                 return addr;
 253         }
 254
 255         if (addr) {
 256                 addr = ALIGN(addr, huge_page_size(h));
 257                 vma = find_vma(mm, addr);
 258                 if (mmap_end - len >= addr &&
 259                     (!vma || addr + len <= vm_start_gap(vma)))
 260                         return addr;
 261         }
 262
 263         /*
 264          * Use mm->get_unmapped_area value as a hint to use topdown routine.
 265          * If architectures have special needs, they should define their own
 266          * version of hugetlb_get_unmapped_area.
 267          */
 268         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
 269                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 270                                 pgoff, flags);
 271         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 272                         pgoff, flags);
 273 }
 274
 275 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 276 static unsigned long
 277 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 278                           unsigned long len, unsigned long pgoff,
 279                           unsigned long flags)
 280 {
 281         return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
 282 }
 283 #endif
 284
 285 /*
 286  * Someone wants to read @bytes from a HWPOISON hugetlb @page from @offset.
 287  * Returns the maximum number of bytes one can read without touching the 1st raw
 288  * HWPOISON subpage.
 289  *
 290  * The implementation borrows the iteration logic from copy_page_to_iter*.
 291  */
 292 static size_t adjust_range_hwpoison(struct page *page, size_t offset, size_t bytes)
 293 {
 294         size_t n = 0;
 295         size_t res = 0;
 296
 297         /* First subpage to start the loop. */
 298         page = nth_page(page, offset / PAGE_SIZE);
 299         offset %= PAGE_SIZE;
 300         while (1) {
 301                 if (is_raw_hwpoison_page_in_hugepage(page))
 302                         break;
 303
 304                 /* Safe to read n bytes without touching HWPOISON subpage. */
 305                 n = min(bytes, (size_t)PAGE_SIZE - offset);
 306                 res += n;
 307                 bytes -= n;
 308                 if (!bytes || !n)
 309                         break;
 310                 offset += n;
 311                 if (offset == PAGE_SIZE) {
 312                         page = nth_page(page, 1);
 313                         offset = 0;
 314                 }
 315         }
 316
 317         return res;
 318 }
 319
 320 /*
 321  * Support for read() - Find the page attached to f_mapping and copy out the
 322  * data. This provides functionality similar to filemap_read().
 323  */
 324 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 325 {
 326         struct file *file = iocb->ki_filp;
 327         struct hstate *h = hstate_file(file);
 328         struct address_space *mapping = file->f_mapping;
 329         struct inode *inode = mapping->host;
 330         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 331         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 332         unsigned long end_index;
 333         loff_t isize;
 334         ssize_t retval = 0;
 335
 336         while (iov_iter_count(to)) {
 337                 struct folio *folio;
 338                 size_t nr, copied, want;
 339
 340                 /* nr is the maximum number of bytes to copy from this page */
 341                 nr = huge_page_size(h);
 342                 isize = i_size_read(inode);
 343                 if (!isize)
 344                         break;
 345                 end_index = (isize - 1) >> huge_page_shift(h);
 346                 if (index > end_index)
 347                         break;
 348                 if (index == end_index) {
 349                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 350                         if (nr <= offset)
 351                                 break;
 352                 }
 353                 nr = nr - offset;
 354
 355                 /* Find the folio */
 356                 folio = filemap_lock_hugetlb_folio(h, mapping, index);
 357                 if (IS_ERR(folio)) {
 358                         /*
 359                          * We have a HOLE, zero out the user-buffer for the
 360                          * length of the hole or request.
 361                          */
 362                         copied = iov_iter_zero(nr, to);
 363                 } else {
 364                         folio_unlock(folio);
 365
 366                         if (!folio_test_has_hwpoisoned(folio))
 367                                 want = nr;
 368                         else {
 369                                 /*
 370                                  * Adjust how many bytes safe to read without
 371                                  * touching the 1st raw HWPOISON subpage after
 372                                  * offset.
 373                                  */
 374                                 want = adjust_range_hwpoison(&folio->page, offset, nr);
 375                                 if (want == 0) {
 376                                         folio_put(folio);
 377                                         retval = -EIO;
 378                                         break;
 379                                 }
 380                         }
 381
 382                         /*
 383                          * We have the folio, copy it to user space buffer.
 384                          */
 385                         copied = copy_folio_to_iter(folio, offset, want, to);
 386                         folio_put(folio);
 387                 }
 388                 offset += copied;
 389                 retval += copied;
 390                 if (copied != nr && iov_iter_count(to)) {
 391                         if (!retval)
 392                                 retval = -EFAULT;
 393                         break;
 394                 }
 395                 index += offset >> huge_page_shift(h);
 396                 offset &= ~huge_page_mask(h);
 397         }
 398         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 399         return retval;
 400 }
 401
 402 static int hugetlbfs_write_begin(struct file *file,
 403                         struct address_space *mapping,
 404                         loff_t pos, unsigned len,
 405                         struct page **pagep, void **fsdata)
 406 {
 407         return -EINVAL;
 408 }
 409
 410 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 411                         loff_t pos, unsigned len, unsigned copied,
 412                         struct page *page, void *fsdata)
 413 {
 414         BUG();
 415         return -EINVAL;
 416 }
 417
 418 static void hugetlb_delete_from_page_cache(struct folio *folio)
 419 {
 420         folio_clear_dirty(folio);
 421         folio_clear_uptodate(folio);
 422         filemap_remove_folio(folio);
 423 }
 424
 425 /*
 426  * Called with i_mmap_rwsem held for inode based vma maps.  This makes
 427  * sure vma (and vm_mm) will not go away.  We also hold the hugetlb fault
 428  * mutex for the page in the mapping.  So, we can not race with page being
 429  * faulted into the vma.
 430  */
 431 static bool hugetlb_vma_maps_page(struct vm_area_struct *vma,
 432                                 unsigned long addr, struct page *page)
 433 {
 434         pte_t *ptep, pte;
 435
 436         ptep = hugetlb_walk(vma, addr, huge_page_size(hstate_vma(vma)));
 437         if (!ptep)
 438                 return false;
 439
 440         pte = huge_ptep_get(ptep);
 441         if (huge_pte_none(pte) || !pte_present(pte))
 442                 return false;
 443
 444         if (pte_page(pte) == page)
 445                 return true;
 446
 447         return false;
 448 }
 449
 450 /*
 451  * Can vma_offset_start/vma_offset_end overflow on 32-bit arches?
 452  * No, because the interval tree returns us only those vmas
 453  * which overlap the truncated area starting at pgoff,
 454  * and no vma on a 32-bit arch can span beyond the 4GB.
 455  */
 456 static unsigned long vma_offset_start(struct vm_area_struct *vma, pgoff_t start)
 457 {
 458         unsigned long offset = 0;
 459
 460         if (vma->vm_pgoff < start)
 461                 offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 462
 463         return vma->vm_start + offset;
 464 }
 465
 466 static unsigned long vma_offset_end(struct vm_area_struct *vma, pgoff_t end)
 467 {
 468         unsigned long t_end;
 469
 470         if (!end)
 471                 return vma->vm_end;
 472
 473         t_end = ((end - vma->vm_pgoff) << PAGE_SHIFT) + vma->vm_start;
 474         if (t_end > vma->vm_end)
 475                 t_end = vma->vm_end;
 476         return t_end;
 477 }
 478
 479 /*
 480  * Called with hugetlb fault mutex held.  Therefore, no more mappings to
 481  * this folio can be created while executing the routine.
 482  */
 483 static void hugetlb_unmap_file_folio(struct hstate *h,
 484                                         struct address_space *mapping,
 485                                         struct folio *folio, pgoff_t index)
 486 {
 487         struct rb_root_cached *root = &mapping->i_mmap;
 488         struct hugetlb_vma_lock *vma_lock;
 489         struct page *page = &folio->page;
 490         struct vm_area_struct *vma;
 491         unsigned long v_start;
 492         unsigned long v_end;
 493         pgoff_t start, end;
 494
 495         start = index * pages_per_huge_page(h);
 496         end = (index + 1) * pages_per_huge_page(h);
 497
 498         i_mmap_lock_write(mapping);
 499 retry:
 500         vma_lock = NULL;
 501         vma_interval_tree_foreach(vma, root, start, end - 1) {
 502                 v_start = vma_offset_start(vma, start);
 503                 v_end = vma_offset_end(vma, end);
 504
 505                 if (!hugetlb_vma_maps_page(vma, v_start, page))
 506                         continue;
 507
 508                 if (!hugetlb_vma_trylock_write(vma)) {
 509                         vma_lock = vma->vm_private_data;
 510                         /*
 511                          * If we can not get vma lock, we need to drop
 512                          * immap_sema and take locks in order.  First,
 513                          * take a ref on the vma_lock structure so that
 514                          * we can be guaranteed it will not go away when
 515                          * dropping immap_sema.
 516                          */
 517                         kref_get(&vma_lock->refs);
 518                         break;
 519                 }
 520
 521                 unmap_hugepage_range(vma, v_start, v_end, NULL,
 522                                      ZAP_FLAG_DROP_MARKER);
 523                 hugetlb_vma_unlock_write(vma);
 524         }
 525
 526         i_mmap_unlock_write(mapping);
 527
 528         if (vma_lock) {
 529                 /*
 530                  * Wait on vma_lock.  We know it is still valid as we have
 531                  * a reference.  We must 'open code' vma locking as we do
 532                  * not know if vma_lock is still attached to vma.
 533                  */
 534                 down_write(&vma_lock->rw_sema);
 535                 i_mmap_lock_write(mapping);
 536
 537                 vma = vma_lock->vma;
 538                 if (!vma) {
 539                         /*
 540                          * If lock is no longer attached to vma, then just
 541                          * unlock, drop our reference and retry looking for
 542                          * other vmas.
 543                          */
 544                         up_write(&vma_lock->rw_sema);
 545                         kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
 546                         goto retry;
 547                 }
 548
 549                 /*
 550                  * vma_lock is still attached to vma.  Check to see if vma
 551                  * still maps page and if so, unmap.
 552                  */
 553                 v_start = vma_offset_start(vma, start);
 554                 v_end = vma_offset_end(vma, end);
 555                 if (hugetlb_vma_maps_page(vma, v_start, page))
 556                         unmap_hugepage_range(vma, v_start, v_end, NULL,
 557                                              ZAP_FLAG_DROP_MARKER);
 558
 559                 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
 560                 hugetlb_vma_unlock_write(vma);
 561
 562                 goto retry;
 563         }
 564 }
 565
 566 static void
 567 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
 568                       zap_flags_t zap_flags)
 569 {
 570         struct vm_area_struct *vma;
 571
 572         /*
 573          * end == 0 indicates that the entire range after start should be
 574          * unmapped.  Note, end is exclusive, whereas the interval tree takes
 575          * an inclusive "last".
 576          */
 577         vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
 578                 unsigned long v_start;
 579                 unsigned long v_end;
 580
 581                 if (!hugetlb_vma_trylock_write(vma))
 582                         continue;
 583
 584                 v_start = vma_offset_start(vma, start);
 585                 v_end = vma_offset_end(vma, end);
 586
 587                 unmap_hugepage_range(vma, v_start, v_end, NULL, zap_flags);
 588
 589                 /*
 590                  * Note that vma lock only exists for shared/non-private
 591                  * vmas.  Therefore, lock is not held when calling
 592                  * unmap_hugepage_range for private vmas.
 593                  */
 594                 hugetlb_vma_unlock_write(vma);
 595         }
 596 }
 597
 598 /*
 599  * Called with hugetlb fault mutex held.
 600  * Returns true if page was actually removed, false otherwise.
 601  */
 602 static bool remove_inode_single_folio(struct hstate *h, struct inode *inode,
 603                                         struct address_space *mapping,
 604                                         struct folio *folio, pgoff_t index,
 605                                         bool truncate_op)
 606 {
 607         bool ret = false;
 608
 609         /*
 610          * If folio is mapped, it was faulted in after being
 611          * unmapped in caller.  Unmap (again) while holding
 612          * the fault mutex.  The mutex will prevent faults
 613          * until we finish removing the folio.
 614          */
 615         if (unlikely(folio_mapped(folio)))
 616                 hugetlb_unmap_file_folio(h, mapping, folio, index);
 617
 618         folio_lock(folio);
 619         /*
 620          * We must remove the folio from page cache before removing
 621          * the region/ reserve map (hugetlb_unreserve_pages).  In
 622          * rare out of memory conditions, removal of the region/reserve
 623          * map could fail.  Correspondingly, the subpool and global
 624          * reserve usage count can need to be adjusted.
 625          */
 626         VM_BUG_ON_FOLIO(folio_test_hugetlb_restore_reserve(folio), folio);
 627         hugetlb_delete_from_page_cache(folio);
 628         ret = true;
 629         if (!truncate_op) {
 630                 if (unlikely(hugetlb_unreserve_pages(inode, index,
 631                                                         index + 1, 1)))
 632                         hugetlb_fix_reserve_counts(inode);
 633         }
 634
 635         folio_unlock(folio);
 636         return ret;
 637 }
 638
 639 /*
 640  * remove_inode_hugepages handles two distinct cases: truncation and hole
 641  * punch.  There are subtle differences in operation for each case.
 642  *
 643  * truncation is indicated by end of range being LLONG_MAX
 644  *      In this case, we first scan the range and release found pages.
 645  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
 646  *      maps and global counts.  Page faults can race with truncation.
 647  *      During faults, hugetlb_no_page() checks i_size before page allocation,
 648  *      and again after obtaining page table lock.  It will 'back out'
 649  *      allocations in the truncated range.
 650  * hole punch is indicated if end is not LLONG_MAX
 651  *      In the hole punch case we scan the range and release found pages.
 652  *      Only when releasing a page is the associated region/reserve map
 653  *      deleted.  The region/reserve map for ranges without associated
 654  *      pages are not modified.  Page faults can race with hole punch.
 655  *      This is indicated if we find a mapped page.
 656  * Note: If the passed end of range value is beyond the end of file, but
 657  * not LLONG_MAX this routine still performs a hole punch operation.
 658  */
 659 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 660                                    loff_t lend)
 661 {
 662         struct hstate *h = hstate_inode(inode);
 663         struct address_space *mapping = &inode->i_data;
 664         const pgoff_t end = lend >> PAGE_SHIFT;
 665         struct folio_batch fbatch;
 666         pgoff_t next, index;
 667         int i, freed = 0;
 668         bool truncate_op = (lend == LLONG_MAX);
 669
 670         folio_batch_init(&fbatch);
 671         next = lstart >> PAGE_SHIFT;
 672         while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
 673                 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
 674                         struct folio *folio = fbatch.folios[i];
 675                         u32 hash = 0;
 676
 677                         index = folio->index >> huge_page_order(h);
 678                         hash = hugetlb_fault_mutex_hash(mapping, index);
 679                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 680
 681                         /*
 682                          * Remove folio that was part of folio_batch.
 683                          */
 684                         if (remove_inode_single_folio(h, inode, mapping, folio,
 685                                                         index, truncate_op))
 686                                 freed++;
 687
 688                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 689                 }
 690                 folio_batch_release(&fbatch);
 691                 cond_resched();
 692         }
 693
 694         if (truncate_op)
 695                 (void)hugetlb_unreserve_pages(inode,
 696                                 lstart >> huge_page_shift(h),
 697                                 LONG_MAX, freed);
 698 }
 699
 700 static void hugetlbfs_evict_inode(struct inode *inode)
 701 {
 702         struct resv_map *resv_map;
 703
 704         remove_inode_hugepages(inode, 0, LLONG_MAX);
 705
 706         /*
 707          * Get the resv_map from the address space embedded in the inode.
 708          * This is the address space which points to any resv_map allocated
 709          * at inode creation time.  If this is a device special inode,
 710          * i_mapping may not point to the original address space.
 711          */
 712         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
 713         /* Only regular and link inodes have associated reserve maps */
 714         if (resv_map)
 715                 resv_map_release(&resv_map->refs);
 716         clear_inode(inode);
 717 }
 718
 719 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 720 {
 721         pgoff_t pgoff;
 722         struct address_space *mapping = inode->i_mapping;
 723         struct hstate *h = hstate_inode(inode);
 724
 725         BUG_ON(offset & ~huge_page_mask(h));
 726         pgoff = offset >> PAGE_SHIFT;
 727
 728         i_size_write(inode, offset);
 729         i_mmap_lock_write(mapping);
 730         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 731                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
 732                                       ZAP_FLAG_DROP_MARKER);
 733         i_mmap_unlock_write(mapping);
 734         remove_inode_hugepages(inode, offset, LLONG_MAX);
 735 }
 736
 737 static void hugetlbfs_zero_partial_page(struct hstate *h,
 738                                         struct address_space *mapping,
 739                                         loff_t start,
 740                                         loff_t end)
 741 {
 742         pgoff_t idx = start >> huge_page_shift(h);
 743         struct folio *folio;
 744
 745         folio = filemap_lock_hugetlb_folio(h, mapping, idx);
 746         if (IS_ERR(folio))
 747                 return;
 748
 749         start = start & ~huge_page_mask(h);
 750         end = end & ~huge_page_mask(h);
 751         if (!end)
 752                 end = huge_page_size(h);
 753
 754         folio_zero_segment(folio, (size_t)start, (size_t)end);
 755
 756         folio_unlock(folio);
 757         folio_put(folio);
 758 }
 759
 760 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 761 {
 762         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 763         struct address_space *mapping = inode->i_mapping;
 764         struct hstate *h = hstate_inode(inode);
 765         loff_t hpage_size = huge_page_size(h);
 766         loff_t hole_start, hole_end;
 767
 768         /*
 769          * hole_start and hole_end indicate the full pages within the hole.
 770          */
 771         hole_start = round_up(offset, hpage_size);
 772         hole_end = round_down(offset + len, hpage_size);
 773
 774         inode_lock(inode);
 775
 776         /* protected by i_rwsem */
 777         if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 778                 inode_unlock(inode);
 779                 return -EPERM;
 780         }
 781
 782         i_mmap_lock_write(mapping);
 783
 784         /* If range starts before first full page, zero partial page. */
 785         if (offset < hole_start)
 786                 hugetlbfs_zero_partial_page(h, mapping,
 787                                 offset, min(offset + len, hole_start));
 788
 789         /* Unmap users of full pages in the hole. */
 790         if (hole_end > hole_start) {
 791                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 792                         hugetlb_vmdelete_list(&mapping->i_mmap,
 793                                               hole_start >> PAGE_SHIFT,
 794                                               hole_end >> PAGE_SHIFT, 0);
 795         }
 796
 797         /* If range extends beyond last full page, zero partial page. */
 798         if ((offset + len) > hole_end && (offset + len) > hole_start)
 799                 hugetlbfs_zero_partial_page(h, mapping,
 800                                 hole_end, offset + len);
 801
 802         i_mmap_unlock_write(mapping);
 803
 804         /* Remove full pages from the file. */
 805         if (hole_end > hole_start)
 806                 remove_inode_hugepages(inode, hole_start, hole_end);
 807
 808         inode_unlock(inode);
 809
 810         return 0;
 811 }
 812
 813 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 814                                 loff_t len)
 815 {
 816         struct inode *inode = file_inode(file);
 817         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 818         struct address_space *mapping = inode->i_mapping;
 819         struct hstate *h = hstate_inode(inode);
 820         struct vm_area_struct pseudo_vma;
 821         struct mm_struct *mm = current->mm;
 822         loff_t hpage_size = huge_page_size(h);
 823         unsigned long hpage_shift = huge_page_shift(h);
 824         pgoff_t start, index, end;
 825         int error;
 826         u32 hash;
 827
 828         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 829                 return -EOPNOTSUPP;
 830
 831         if (mode & FALLOC_FL_PUNCH_HOLE)
 832                 return hugetlbfs_punch_hole(inode, offset, len);
 833
 834         /*
 835          * Default preallocate case.
 836          * For this range, start is rounded down and end is rounded up
 837          * as well as being converted to page offsets.
 838          */
 839         start = offset >> hpage_shift;
 840         end = (offset + len + hpage_size - 1) >> hpage_shift;
 841
 842         inode_lock(inode);
 843
 844         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 845         error = inode_newsize_ok(inode, offset + len);
 846         if (error)
 847                 goto out;
 848
 849         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 850                 error = -EPERM;
 851                 goto out;
 852         }
 853
 854         /*
 855          * Initialize a pseudo vma as this is required by the huge page
 856          * allocation routines.  If NUMA is configured, use page index
 857          * as input to create an allocation policy.
 858          */
 859         vma_init(&pseudo_vma, mm);
 860         vm_flags_init(&pseudo_vma, VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 861         pseudo_vma.vm_file = file;
 862
 863         for (index = start; index < end; index++) {
 864                 /*
 865                  * This is supposed to be the vaddr where the page is being
 866                  * faulted in, but we have no vaddr here.
 867                  */
 868                 struct folio *folio;
 869                 unsigned long addr;
 870
 871                 cond_resched();
 872
 873                 /*
 874                  * fallocate(2) manpage permits EINTR; we may have been
 875                  * interrupted because we are using up too much memory.
 876                  */
 877                 if (signal_pending(current)) {
 878                         error = -EINTR;
 879                         break;
 880                 }
 881
 882                 /* addr is the offset within the file (zero based) */
 883                 addr = index * hpage_size;
 884
 885                 /* mutex taken here, fault path and hole punch */
 886                 hash = hugetlb_fault_mutex_hash(mapping, index);
 887                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 888
 889                 /* See if already present in mapping to avoid alloc/free */
 890                 folio = filemap_get_folio(mapping, index << huge_page_order(h));
 891                 if (!IS_ERR(folio)) {
 892                         folio_put(folio);
 893                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 894                         continue;
 895                 }
 896
 897                 /*
 898                  * Allocate folio without setting the avoid_reserve argument.
 899                  * There certainly are no reserves associated with the
 900                  * pseudo_vma.  However, there could be shared mappings with
 901                  * reserves for the file at the inode level.  If we fallocate
 902                  * folios in these areas, we need to consume the reserves
 903                  * to keep reservation accounting consistent.
 904                  */
 905                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 906                 folio = alloc_hugetlb_folio(&pseudo_vma, addr, 0);
 907                 hugetlb_drop_vma_policy(&pseudo_vma);
 908                 if (IS_ERR(folio)) {
 909                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 910                         error = PTR_ERR(folio);
 911                         goto out;
 912                 }
 913                 clear_huge_page(&folio->page, addr, pages_per_huge_page(h));
 914                 __folio_mark_uptodate(folio);
 915                 error = hugetlb_add_to_page_cache(folio, mapping, index);
 916                 if (unlikely(error)) {
 917                         restore_reserve_on_error(h, &pseudo_vma, addr, folio);
 918                         folio_put(folio);
 919                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 920                         goto out;
 921                 }
 922
 923                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 924
 925                 folio_set_hugetlb_migratable(folio);
 926                 /*
 927                  * folio_unlock because locked by hugetlb_add_to_page_cache()
 928                  * folio_put() due to reference from alloc_hugetlb_folio()
 929                  */
 930                 folio_unlock(folio);
 931                 folio_put(folio);
 932         }
 933
 934         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 935                 i_size_write(inode, offset + len);
 936         inode_set_ctime_current(inode);
 937 out:
 938         inode_unlock(inode);
 939         return error;
 940 }
 941
 942 static int hugetlbfs_setattr(struct mnt_idmap *idmap,
 943                              struct dentry *dentry, struct iattr *attr)
 944 {
 945         struct inode *inode = d_inode(dentry);
 946         struct hstate *h = hstate_inode(inode);
 947         int error;
 948         unsigned int ia_valid = attr->ia_valid;
 949         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 950
 951         error = setattr_prepare(&nop_mnt_idmap, dentry, attr);
 952         if (error)
 953                 return error;
 954
 955         if (ia_valid & ATTR_SIZE) {
 956                 loff_t oldsize = inode->i_size;
 957                 loff_t newsize = attr->ia_size;
 958
 959                 if (newsize & ~huge_page_mask(h))
 960                         return -EINVAL;
 961                 /* protected by i_rwsem */
 962                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 963                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 964                         return -EPERM;
 965                 hugetlb_vmtruncate(inode, newsize);
 966         }
 967
 968         setattr_copy(&nop_mnt_idmap, inode, attr);
 969         mark_inode_dirty(inode);
 970         return 0;
 971 }
 972
 973 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 974                                         struct hugetlbfs_fs_context *ctx)
 975 {
 976         struct inode *inode;
 977
 978         inode = new_inode(sb);
 979         if (inode) {
 980                 inode->i_ino = get_next_ino();
 981                 inode->i_mode = S_IFDIR | ctx->mode;
 982                 inode->i_uid = ctx->uid;
 983                 inode->i_gid = ctx->gid;
 984                 inode->i_atime = inode->i_mtime = inode_set_ctime_current(inode);
 985                 inode->i_op = &hugetlbfs_dir_inode_operations;
 986                 inode->i_fop = &simple_dir_operations;
 987                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 988                 inc_nlink(inode);
 989                 lockdep_annotate_inode_mutex_key(inode);
 990         }
 991         return inode;
 992 }
 993
 994 /*
 995  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 996  * be taken from reclaim -- unlike regular filesystems. This needs an
 997  * annotation because huge_pmd_share() does an allocation under hugetlb's
 998  * i_mmap_rwsem.
 999  */
1000 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
1001
1002 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
1003                                         struct inode *dir,
1004                                         umode_t mode, dev_t dev)
1005 {
1006         struct inode *inode;
1007         struct resv_map *resv_map = NULL;
1008
1009         /*
1010          * Reserve maps are only needed for inodes that can have associated
1011          * page allocations.
1012          */
1013         if (S_ISREG(mode) || S_ISLNK(mode)) {
1014                 resv_map = resv_map_alloc();
1015                 if (!resv_map)
1016                         return NULL;
1017         }
1018
1019         inode = new_inode(sb);
1020         if (inode) {
1021                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
1022
1023                 inode->i_ino = get_next_ino();
1024                 inode_init_owner(&nop_mnt_idmap, inode, dir, mode);
1025                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
1026                                 &hugetlbfs_i_mmap_rwsem_key);
1027                 inode->i_mapping->a_ops = &hugetlbfs_aops;
1028                 inode->i_atime = inode->i_mtime = inode_set_ctime_current(inode);
1029                 inode->i_mapping->private_data = resv_map;
1030                 info->seals = F_SEAL_SEAL;
1031                 switch (mode & S_IFMT) {
1032                 default:
1033                         init_special_inode(inode, mode, dev);
1034                         break;
1035                 case S_IFREG:
1036                         inode->i_op = &hugetlbfs_inode_operations;
1037                         inode->i_fop = &hugetlbfs_file_operations;
1038                         break;
1039                 case S_IFDIR:
1040                         inode->i_op = &hugetlbfs_dir_inode_operations;
1041                         inode->i_fop = &simple_dir_operations;
1042
1043                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
1044                         inc_nlink(inode);
1045                         break;
1046                 case S_IFLNK:
1047                         inode->i_op = &page_symlink_inode_operations;
1048                         inode_nohighmem(inode);
1049                         break;
1050                 }
1051                 lockdep_annotate_inode_mutex_key(inode);
1052         } else {
1053                 if (resv_map)
1054                         kref_put(&resv_map->refs, resv_map_release);
1055         }
1056
1057         return inode;
1058 }
1059
1060 /*
1061  * File creation. Allocate an inode, and we're done..
1062  */
1063 static int hugetlbfs_mknod(struct mnt_idmap *idmap, struct inode *dir,
1064                            struct dentry *dentry, umode_t mode, dev_t dev)
1065 {
1066         struct inode *inode;
1067
1068         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
1069         if (!inode)
1070                 return -ENOSPC;
1071         dir->i_mtime = inode_set_ctime_current(dir);
1072         d_instantiate(dentry, inode);
1073         dget(dentry);/* Extra count - pin the dentry in core */
1074         return 0;
1075 }
1076
1077 static int hugetlbfs_mkdir(struct mnt_idmap *idmap, struct inode *dir,
1078                            struct dentry *dentry, umode_t mode)
1079 {
1080         int retval = hugetlbfs_mknod(&nop_mnt_idmap, dir, dentry,
1081                                      mode | S_IFDIR, 0);
1082         if (!retval)
1083                 inc_nlink(dir);
1084         return retval;
1085 }
1086
1087 static int hugetlbfs_create(struct mnt_idmap *idmap,
1088                             struct inode *dir, struct dentry *dentry,
1089                             umode_t mode, bool excl)
1090 {
1091         return hugetlbfs_mknod(&nop_mnt_idmap, dir, dentry, mode | S_IFREG, 0);
1092 }
1093
1094 static int hugetlbfs_tmpfile(struct mnt_idmap *idmap,
1095                              struct inode *dir, struct file *file,
1096                              umode_t mode)
1097 {
1098         struct inode *inode;
1099
1100         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode | S_IFREG, 0);
1101         if (!inode)
1102                 return -ENOSPC;
1103         dir->i_mtime = inode_set_ctime_current(dir);
1104         d_tmpfile(file, inode);
1105         return finish_open_simple(file, 0);
1106 }
1107
1108 static int hugetlbfs_symlink(struct mnt_idmap *idmap,
1109                              struct inode *dir, struct dentry *dentry,
1110                              const char *symname)
1111 {
1112         struct inode *inode;
1113         int error = -ENOSPC;
1114
1115         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
1116         if (inode) {
1117                 int l = strlen(symname)+1;
1118                 error = page_symlink(inode, symname, l);
1119                 if (!error) {
1120                         d_instantiate(dentry, inode);
1121                         dget(dentry);
1122                 } else
1123                         iput(inode);
1124         }
1125         dir->i_mtime = inode_set_ctime_current(dir);
1126
1127         return error;
1128 }
1129
1130 #ifdef CONFIG_MIGRATION
1131 static int hugetlbfs_migrate_folio(struct address_space *mapping,
1132                                 struct folio *dst, struct folio *src,
1133                                 enum migrate_mode mode)
1134 {
1135         int rc;
1136
1137         rc = migrate_huge_page_move_mapping(mapping, dst, src);
1138         if (rc != MIGRATEPAGE_SUCCESS)
1139                 return rc;
1140
1141         if (hugetlb_folio_subpool(src)) {
1142                 hugetlb_set_folio_subpool(dst,
1143                                         hugetlb_folio_subpool(src));
1144                 hugetlb_set_folio_subpool(src, NULL);
1145         }
1146
1147         if (mode != MIGRATE_SYNC_NO_COPY)
1148                 folio_migrate_copy(dst, src);
1149         else
1150                 folio_migrate_flags(dst, src);
1151
1152         return MIGRATEPAGE_SUCCESS;
1153 }
1154 #else
1155 #define hugetlbfs_migrate_folio NULL
1156 #endif
1157
1158 static int hugetlbfs_error_remove_page(struct address_space *mapping,
1159                                 struct page *page)
1160 {
1161         return 0;
1162 }
1163
1164 /*
1165  * Display the mount options in /proc/mounts.
1166  */
1167 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1168 {
1169         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1170         struct hugepage_subpool *spool = sbinfo->spool;
1171         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1172         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1173         char mod;
1174
1175         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1176                 seq_printf(m, ",uid=%u",
1177                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1178         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1179                 seq_printf(m, ",gid=%u",
1180                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1181         if (sbinfo->mode != 0755)
1182                 seq_printf(m, ",mode=%o", sbinfo->mode);
1183         if (sbinfo->max_inodes != -1)
1184                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1185
1186         hpage_size /= 1024;
1187         mod = 'K';
1188         if (hpage_size >= 1024) {
1189                 hpage_size /= 1024;
1190                 mod = 'M';
1191         }
1192         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1193         if (spool) {
1194                 if (spool->max_hpages != -1)
1195                         seq_printf(m, ",size=%llu",
1196                                    (unsigned long long)spool->max_hpages << hpage_shift);
1197                 if (spool->min_hpages != -1)
1198                         seq_printf(m, ",min_size=%llu",
1199                                    (unsigned long long)spool->min_hpages << hpage_shift);
1200         }
1201         return 0;
1202 }
1203
1204 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1205 {
1206         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1207         struct hstate *h = hstate_inode(d_inode(dentry));
1208
1209         buf->f_type = HUGETLBFS_MAGIC;
1210         buf->f_bsize = huge_page_size(h);
1211         if (sbinfo) {
1212                 spin_lock(&sbinfo->stat_lock);
1213                 /* If no limits set, just report 0 or -1 for max/free/used
1214                  * blocks, like simple_statfs() */
1215                 if (sbinfo->spool) {
1216                         long free_pages;
1217
1218                         spin_lock_irq(&sbinfo->spool->lock);
1219                         buf->f_blocks = sbinfo->spool->max_hpages;
1220                         free_pages = sbinfo->spool->max_hpages
1221                                 - sbinfo->spool->used_hpages;
1222                         buf->f_bavail = buf->f_bfree = free_pages;
1223                         spin_unlock_irq(&sbinfo->spool->lock);
1224                         buf->f_files = sbinfo->max_inodes;
1225                         buf->f_ffree = sbinfo->free_inodes;
1226                 }
1227                 spin_unlock(&sbinfo->stat_lock);
1228         }
1229         buf->f_namelen = NAME_MAX;
1230         return 0;
1231 }
1232
1233 static void hugetlbfs_put_super(struct super_block *sb)
1234 {
1235         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1236
1237         if (sbi) {
1238                 sb->s_fs_info = NULL;
1239
1240                 if (sbi->spool)
1241                         hugepage_put_subpool(sbi->spool);
1242
1243                 kfree(sbi);
1244         }
1245 }
1246
1247 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1248 {
1249         if (sbinfo->free_inodes >= 0) {
1250                 spin_lock(&sbinfo->stat_lock);
1251                 if (unlikely(!sbinfo->free_inodes)) {
1252                         spin_unlock(&sbinfo->stat_lock);
1253                         return 0;
1254                 }
1255                 sbinfo->free_inodes--;
1256                 spin_unlock(&sbinfo->stat_lock);
1257         }
1258
1259         return 1;
1260 }
1261
1262 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1263 {
1264         if (sbinfo->free_inodes >= 0) {
1265                 spin_lock(&sbinfo->stat_lock);
1266                 sbinfo->free_inodes++;
1267                 spin_unlock(&sbinfo->stat_lock);
1268         }
1269 }
1270
1271
1272 static struct kmem_cache *hugetlbfs_inode_cachep;
1273
1274 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1275 {
1276         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1277         struct hugetlbfs_inode_info *p;
1278
1279         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1280                 return NULL;
1281         p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1282         if (unlikely(!p)) {
1283                 hugetlbfs_inc_free_inodes(sbinfo);
1284                 return NULL;
1285         }
1286
1287         /*
1288          * Any time after allocation, hugetlbfs_destroy_inode can be called
1289          * for the inode.  mpol_free_shared_policy is unconditionally called
1290          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1291          * in case of a quick call to destroy.
1292          *
1293          * Note that the policy is initialized even if we are creating a
1294          * private inode.  This simplifies hugetlbfs_destroy_inode.
1295          */
1296         mpol_shared_policy_init(&p->policy, NULL);
1297
1298         return &p->vfs_inode;
1299 }
1300
1301 static void hugetlbfs_free_inode(struct inode *inode)
1302 {
1303         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1304 }
1305
1306 static void hugetlbfs_destroy_inode(struct inode *inode)
1307 {
1308         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1309         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1310 }
1311
1312 static const struct address_space_operations hugetlbfs_aops = {
1313         .write_begin    = hugetlbfs_write_begin,
1314         .write_end      = hugetlbfs_write_end,
1315         .dirty_folio    = noop_dirty_folio,
1316         .migrate_folio  = hugetlbfs_migrate_folio,
1317         .error_remove_page      = hugetlbfs_error_remove_page,
1318 };
1319
1320
1321 static void init_once(void *foo)
1322 {
1323         struct hugetlbfs_inode_info *ei = foo;
1324
1325         inode_init_once(&ei->vfs_inode);
1326 }
1327
1328 const struct file_operations hugetlbfs_file_operations = {
1329         .read_iter              = hugetlbfs_read_iter,
1330         .mmap                   = hugetlbfs_file_mmap,
1331         .fsync                  = noop_fsync,
1332         .get_unmapped_area      = hugetlb_get_unmapped_area,
1333         .llseek                 = default_llseek,
1334         .fallocate              = hugetlbfs_fallocate,
1335 };
1336
1337 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1338         .create         = hugetlbfs_create,
1339         .lookup         = simple_lookup,
1340         .link           = simple_link,
1341         .unlink         = simple_unlink,
1342         .symlink        = hugetlbfs_symlink,
1343         .mkdir          = hugetlbfs_mkdir,
1344         .rmdir          = simple_rmdir,
1345         .mknod          = hugetlbfs_mknod,
1346         .rename         = simple_rename,
1347         .setattr        = hugetlbfs_setattr,
1348         .tmpfile        = hugetlbfs_tmpfile,
1349 };
1350
1351 static const struct inode_operations hugetlbfs_inode_operations = {
1352         .setattr        = hugetlbfs_setattr,
1353 };
1354
1355 static const struct super_operations hugetlbfs_ops = {
1356         .alloc_inode    = hugetlbfs_alloc_inode,
1357         .free_inode     = hugetlbfs_free_inode,
1358         .destroy_inode  = hugetlbfs_destroy_inode,
1359         .evict_inode    = hugetlbfs_evict_inode,
1360         .statfs         = hugetlbfs_statfs,
1361         .put_super      = hugetlbfs_put_super,
1362         .show_options   = hugetlbfs_show_options,
1363 };
1364
1365 /*
1366  * Convert size option passed from command line to number of huge pages
1367  * in the pool specified by hstate.  Size option could be in bytes
1368  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1369  */
1370 static long
1371 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1372                          enum hugetlbfs_size_type val_type)
1373 {
1374         if (val_type == NO_SIZE)
1375                 return -1;
1376
1377         if (val_type == SIZE_PERCENT) {
1378                 size_opt <<= huge_page_shift(h);
1379                 size_opt *= h->max_huge_pages;
1380                 do_div(size_opt, 100);
1381         }
1382
1383         size_opt >>= huge_page_shift(h);
1384         return size_opt;
1385 }
1386
1387 /*
1388  * Parse one mount parameter.
1389  */
1390 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1391 {
1392         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1393         struct fs_parse_result result;
1394         char *rest;
1395         unsigned long ps;
1396         int opt;
1397
1398         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1399         if (opt < 0)
1400                 return opt;
1401
1402         switch (opt) {
1403         case Opt_uid:
1404                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1405                 if (!uid_valid(ctx->uid))
1406                         goto bad_val;
1407                 return 0;
1408
1409         case Opt_gid:
1410                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1411                 if (!gid_valid(ctx->gid))
1412                         goto bad_val;
1413                 return 0;
1414
1415         case Opt_mode:
1416                 ctx->mode = result.uint_32 & 01777U;
1417                 return 0;
1418
1419         case Opt_size:
1420                 /* memparse() will accept a K/M/G without a digit */
1421                 if (!param->string || !isdigit(param->string[0]))
1422                         goto bad_val;
1423                 ctx->max_size_opt = memparse(param->string, &rest);
1424                 ctx->max_val_type = SIZE_STD;
1425                 if (*rest == '%')
1426                         ctx->max_val_type = SIZE_PERCENT;
1427                 return 0;
1428
1429         case Opt_nr_inodes:
1430                 /* memparse() will accept a K/M/G without a digit */
1431                 if (!param->string || !isdigit(param->string[0]))
1432                         goto bad_val;
1433                 ctx->nr_inodes = memparse(param->string, &rest);
1434                 return 0;
1435
1436         case Opt_pagesize:
1437                 ps = memparse(param->string, &rest);
1438                 ctx->hstate = size_to_hstate(ps);
1439                 if (!ctx->hstate) {
1440                         pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1441                         return -EINVAL;
1442                 }
1443                 return 0;
1444
1445         case Opt_min_size:
1446                 /* memparse() will accept a K/M/G without a digit */
1447                 if (!param->string || !isdigit(param->string[0]))
1448                         goto bad_val;
1449                 ctx->min_size_opt = memparse(param->string, &rest);
1450                 ctx->min_val_type = SIZE_STD;
1451                 if (*rest == '%')
1452                         ctx->min_val_type = SIZE_PERCENT;
1453                 return 0;
1454
1455         default:
1456                 return -EINVAL;
1457         }
1458
1459 bad_val:
1460         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1461                       param->string, param->key);
1462 }
1463
1464 /*
1465  * Validate the parsed options.
1466  */
1467 static int hugetlbfs_validate(struct fs_context *fc)
1468 {
1469         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1470
1471         /*
1472          * Use huge page pool size (in hstate) to convert the size
1473          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1474          */
1475         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1476                                                    ctx->max_size_opt,
1477                                                    ctx->max_val_type);
1478         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1479                                                    ctx->min_size_opt,
1480                                                    ctx->min_val_type);
1481
1482         /*
1483          * If max_size was specified, then min_size must be smaller
1484          */
1485         if (ctx->max_val_type > NO_SIZE &&
1486             ctx->min_hpages > ctx->max_hpages) {
1487                 pr_err("Minimum size can not be greater than maximum size\n");
1488                 return -EINVAL;
1489         }
1490
1491         return 0;
1492 }
1493
1494 static int
1495 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1496 {
1497         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1498         struct hugetlbfs_sb_info *sbinfo;
1499
1500         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1501         if (!sbinfo)
1502                 return -ENOMEM;
1503         sb->s_fs_info = sbinfo;
1504         spin_lock_init(&sbinfo->stat_lock);
1505         sbinfo->hstate          = ctx->hstate;
1506         sbinfo->max_inodes      = ctx->nr_inodes;
1507         sbinfo->free_inodes     = ctx->nr_inodes;
1508         sbinfo->spool           = NULL;
1509         sbinfo->uid             = ctx->uid;
1510         sbinfo->gid             = ctx->gid;
1511         sbinfo->mode            = ctx->mode;
1512
1513         /*
1514          * Allocate and initialize subpool if maximum or minimum size is
1515          * specified.  Any needed reservations (for minimum size) are taken
1516          * when the subpool is created.
1517          */
1518         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1519                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1520                                                      ctx->max_hpages,
1521                                                      ctx->min_hpages);
1522                 if (!sbinfo->spool)
1523                         goto out_free;
1524         }
1525         sb->s_maxbytes = MAX_LFS_FILESIZE;
1526         sb->s_blocksize = huge_page_size(ctx->hstate);
1527         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1528         sb->s_magic = HUGETLBFS_MAGIC;
1529         sb->s_op = &hugetlbfs_ops;
1530         sb->s_time_gran = 1;
1531
1532         /*
1533          * Due to the special and limited functionality of hugetlbfs, it does
1534          * not work well as a stacking filesystem.
1535          */
1536         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1537         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1538         if (!sb->s_root)
1539                 goto out_free;
1540         return 0;
1541 out_free:
1542         kfree(sbinfo->spool);
1543         kfree(sbinfo);
1544         return -ENOMEM;
1545 }
1546
1547 static int hugetlbfs_get_tree(struct fs_context *fc)
1548 {
1549         int err = hugetlbfs_validate(fc);
1550         if (err)
1551                 return err;
1552         return get_tree_nodev(fc, hugetlbfs_fill_super);
1553 }
1554
1555 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1556 {
1557         kfree(fc->fs_private);
1558 }
1559
1560 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1561         .free           = hugetlbfs_fs_context_free,
1562         .parse_param    = hugetlbfs_parse_param,
1563         .get_tree       = hugetlbfs_get_tree,
1564 };
1565
1566 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1567 {
1568         struct hugetlbfs_fs_context *ctx;
1569
1570         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1571         if (!ctx)
1572                 return -ENOMEM;
1573
1574         ctx->max_hpages = -1; /* No limit on size by default */
1575         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1576         ctx->uid        = current_fsuid();
1577         ctx->gid        = current_fsgid();
1578         ctx->mode       = 0755;
1579         ctx->hstate     = &default_hstate;
1580         ctx->min_hpages = -1; /* No default minimum size */
1581         ctx->max_val_type = NO_SIZE;
1582         ctx->min_val_type = NO_SIZE;
1583         fc->fs_private = ctx;
1584         fc->ops = &hugetlbfs_fs_context_ops;
1585         return 0;
1586 }
1587
1588 static struct file_system_type hugetlbfs_fs_type = {
1589         .name                   = "hugetlbfs",
1590         .init_fs_context        = hugetlbfs_init_fs_context,
1591         .parameters             = hugetlb_fs_parameters,
1592         .kill_sb                = kill_litter_super,
1593 };
1594
1595 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1596
1597 static int can_do_hugetlb_shm(void)
1598 {
1599         kgid_t shm_group;
1600         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1601         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1602 }
1603
1604 static int get_hstate_idx(int page_size_log)
1605 {
1606         struct hstate *h = hstate_sizelog(page_size_log);
1607
1608         if (!h)
1609                 return -1;
1610         return hstate_index(h);
1611 }
1612
1613 /*
1614  * Note that size should be aligned to proper hugepage size in caller side,
1615  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1616  */
1617 struct file *hugetlb_file_setup(const char *name, size_t size,
1618                                 vm_flags_t acctflag, int creat_flags,
1619                                 int page_size_log)
1620 {
1621         struct inode *inode;
1622         struct vfsmount *mnt;
1623         int hstate_idx;
1624         struct file *file;
1625
1626         hstate_idx = get_hstate_idx(page_size_log);
1627         if (hstate_idx < 0)
1628                 return ERR_PTR(-ENODEV);
1629
1630         mnt = hugetlbfs_vfsmount[hstate_idx];
1631         if (!mnt)
1632                 return ERR_PTR(-ENOENT);
1633
1634         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1635                 struct ucounts *ucounts = current_ucounts();
1636
1637                 if (user_shm_lock(size, ucounts)) {
1638                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1639                                 current->comm, current->pid);
1640                         user_shm_unlock(size, ucounts);
1641                 }
1642                 return ERR_PTR(-EPERM);
1643         }
1644
1645         file = ERR_PTR(-ENOSPC);
1646         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1647         if (!inode)
1648                 goto out;
1649         if (creat_flags == HUGETLB_SHMFS_INODE)
1650                 inode->i_flags |= S_PRIVATE;
1651
1652         inode->i_size = size;
1653         clear_nlink(inode);
1654
1655         if (!hugetlb_reserve_pages(inode, 0,
1656                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1657                         acctflag))
1658                 file = ERR_PTR(-ENOMEM);
1659         else
1660                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1661                                         &hugetlbfs_file_operations);
1662         if (!IS_ERR(file))
1663                 return file;
1664
1665         iput(inode);
1666 out:
1667         return file;
1668 }
1669
1670 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1671 {
1672         struct fs_context *fc;
1673         struct vfsmount *mnt;
1674
1675         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1676         if (IS_ERR(fc)) {
1677                 mnt = ERR_CAST(fc);
1678         } else {
1679                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1680                 ctx->hstate = h;
1681                 mnt = fc_mount(fc);
1682                 put_fs_context(fc);
1683         }
1684         if (IS_ERR(mnt))
1685                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1686                        huge_page_size(h) / SZ_1K);
1687         return mnt;
1688 }
1689
1690 static int __init init_hugetlbfs_fs(void)
1691 {
1692         struct vfsmount *mnt;
1693         struct hstate *h;
1694         int error;
1695         int i;
1696
1697         if (!hugepages_supported()) {
1698                 pr_info("disabling because there are no supported hugepage sizes\n");
1699                 return -ENOTSUPP;
1700         }
1701
1702         error = -ENOMEM;
1703         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1704                                         sizeof(struct hugetlbfs_inode_info),
1705                                         0, SLAB_ACCOUNT, init_once);
1706         if (hugetlbfs_inode_cachep == NULL)
1707                 goto out;
1708
1709         error = register_filesystem(&hugetlbfs_fs_type);
1710         if (error)
1711                 goto out_free;
1712
1713         /* default hstate mount is required */
1714         mnt = mount_one_hugetlbfs(&default_hstate);
1715         if (IS_ERR(mnt)) {
1716                 error = PTR_ERR(mnt);
1717                 goto out_unreg;
1718         }
1719         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1720
1721         /* other hstates are optional */
1722         i = 0;
1723         for_each_hstate(h) {
1724                 if (i == default_hstate_idx) {
1725                         i++;
1726                         continue;
1727                 }
1728
1729                 mnt = mount_one_hugetlbfs(h);
1730                 if (IS_ERR(mnt))
1731                         hugetlbfs_vfsmount[i] = NULL;
1732                 else
1733                         hugetlbfs_vfsmount[i] = mnt;
1734                 i++;
1735         }
1736
1737         return 0;
1738
1739  out_unreg:
1740         (void)unregister_filesystem(&hugetlbfs_fs_type);
1741  out_free:
1742         kmem_cache_destroy(hugetlbfs_inode_cachep);
1743  out:
1744         return error;
1745 }
1746 fs_initcall(init_hugetlbfs_fs)