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 += 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++;
 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 page *page;
 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 page */
 356                 page = find_lock_page(mapping, index);
 357                 if (unlikely(page == NULL)) {
 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                         unlock_page(page);
 365
 366                         if (!PageHWPoison(page))
 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(page, offset, nr);
 375                                 if (want == 0) {
 376                                         put_page(page);
 377                                         retval = -EIO;
 378                                         break;
 379                                 }
 380                         }
 381
 382                         /*
 383                          * We have the page, copy it to user space buffer.
 384                          */
 385                         copied = copy_page_to_iter(page, offset, want, to);
 386                         put_page(page);
 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 start = lstart >> huge_page_shift(h);
 665         const pgoff_t end = lend >> huge_page_shift(h);
 666         struct folio_batch fbatch;
 667         pgoff_t next, index;
 668         int i, freed = 0;
 669         bool truncate_op = (lend == LLONG_MAX);
 670
 671         folio_batch_init(&fbatch);
 672         next = start;
 673         while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
 674                 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
 675                         struct folio *folio = fbatch.folios[i];
 676                         u32 hash = 0;
 677
 678                         index = folio->index;
 679                         hash = hugetlb_fault_mutex_hash(mapping, index);
 680                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 681
 682                         /*
 683                          * Remove folio that was part of folio_batch.
 684                          */
 685                         if (remove_inode_single_folio(h, inode, mapping, folio,
 686                                                         index, truncate_op))
 687                                 freed++;
 688
 689                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 690                 }
 691                 folio_batch_release(&fbatch);
 692                 cond_resched();
 693         }
 694
 695         if (truncate_op)
 696                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 697 }
 698
 699 static void hugetlbfs_evict_inode(struct inode *inode)
 700 {
 701         struct resv_map *resv_map;
 702
 703         remove_inode_hugepages(inode, 0, LLONG_MAX);
 704
 705         /*
 706          * Get the resv_map from the address space embedded in the inode.
 707          * This is the address space which points to any resv_map allocated
 708          * at inode creation time.  If this is a device special inode,
 709          * i_mapping may not point to the original address space.
 710          */
 711         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
 712         /* Only regular and link inodes have associated reserve maps */
 713         if (resv_map)
 714                 resv_map_release(&resv_map->refs);
 715         clear_inode(inode);
 716 }
 717
 718 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 719 {
 720         pgoff_t pgoff;
 721         struct address_space *mapping = inode->i_mapping;
 722         struct hstate *h = hstate_inode(inode);
 723
 724         BUG_ON(offset & ~huge_page_mask(h));
 725         pgoff = offset >> PAGE_SHIFT;
 726
 727         i_size_write(inode, offset);
 728         i_mmap_lock_write(mapping);
 729         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 730                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
 731                                       ZAP_FLAG_DROP_MARKER);
 732         i_mmap_unlock_write(mapping);
 733         remove_inode_hugepages(inode, offset, LLONG_MAX);
 734 }
 735
 736 static void hugetlbfs_zero_partial_page(struct hstate *h,
 737                                         struct address_space *mapping,
 738                                         loff_t start,
 739                                         loff_t end)
 740 {
 741         pgoff_t idx = start >> huge_page_shift(h);
 742         struct folio *folio;
 743
 744         folio = filemap_lock_folio(mapping, idx);
 745         if (IS_ERR(folio))
 746                 return;
 747
 748         start = start & ~huge_page_mask(h);
 749         end = end & ~huge_page_mask(h);
 750         if (!end)
 751                 end = huge_page_size(h);
 752
 753         folio_zero_segment(folio, (size_t)start, (size_t)end);
 754
 755         folio_unlock(folio);
 756         folio_put(folio);
 757 }
 758
 759 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 760 {
 761         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 762         struct address_space *mapping = inode->i_mapping;
 763         struct hstate *h = hstate_inode(inode);
 764         loff_t hpage_size = huge_page_size(h);
 765         loff_t hole_start, hole_end;
 766
 767         /*
 768          * hole_start and hole_end indicate the full pages within the hole.
 769          */
 770         hole_start = round_up(offset, hpage_size);
 771         hole_end = round_down(offset + len, hpage_size);
 772
 773         inode_lock(inode);
 774
 775         /* protected by i_rwsem */
 776         if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 777                 inode_unlock(inode);
 778                 return -EPERM;
 779         }
 780
 781         i_mmap_lock_write(mapping);
 782
 783         /* If range starts before first full page, zero partial page. */
 784         if (offset < hole_start)
 785                 hugetlbfs_zero_partial_page(h, mapping,
 786                                 offset, min(offset + len, hole_start));
 787
 788         /* Unmap users of full pages in the hole. */
 789         if (hole_end > hole_start) {
 790                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 791                         hugetlb_vmdelete_list(&mapping->i_mmap,
 792                                               hole_start >> PAGE_SHIFT,
 793                                               hole_end >> PAGE_SHIFT, 0);
 794         }
 795
 796         /* If range extends beyond last full page, zero partial page. */
 797         if ((offset + len) > hole_end && (offset + len) > hole_start)
 798                 hugetlbfs_zero_partial_page(h, mapping,
 799                                 hole_end, offset + len);
 800
 801         i_mmap_unlock_write(mapping);
 802
 803         /* Remove full pages from the file. */
 804         if (hole_end > hole_start)
 805                 remove_inode_hugepages(inode, hole_start, hole_end);
 806
 807         inode_unlock(inode);
 808
 809         return 0;
 810 }
 811
 812 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 813                                 loff_t len)
 814 {
 815         struct inode *inode = file_inode(file);
 816         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 817         struct address_space *mapping = inode->i_mapping;
 818         struct hstate *h = hstate_inode(inode);
 819         struct vm_area_struct pseudo_vma;
 820         struct mm_struct *mm = current->mm;
 821         loff_t hpage_size = huge_page_size(h);
 822         unsigned long hpage_shift = huge_page_shift(h);
 823         pgoff_t start, index, end;
 824         int error;
 825         u32 hash;
 826
 827         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 828                 return -EOPNOTSUPP;
 829
 830         if (mode & FALLOC_FL_PUNCH_HOLE)
 831                 return hugetlbfs_punch_hole(inode, offset, len);
 832
 833         /*
 834          * Default preallocate case.
 835          * For this range, start is rounded down and end is rounded up
 836          * as well as being converted to page offsets.
 837          */
 838         start = offset >> hpage_shift;
 839         end = (offset + len + hpage_size - 1) >> hpage_shift;
 840
 841         inode_lock(inode);
 842
 843         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 844         error = inode_newsize_ok(inode, offset + len);
 845         if (error)
 846                 goto out;
 847
 848         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 849                 error = -EPERM;
 850                 goto out;
 851         }
 852
 853         /*
 854          * Initialize a pseudo vma as this is required by the huge page
 855          * allocation routines.  If NUMA is configured, use page index
 856          * as input to create an allocation policy.
 857          */
 858         vma_init(&pseudo_vma, mm);
 859         vm_flags_init(&pseudo_vma, VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 860         pseudo_vma.vm_file = file;
 861
 862         for (index = start; index < end; index++) {
 863                 /*
 864                  * This is supposed to be the vaddr where the page is being
 865                  * faulted in, but we have no vaddr here.
 866                  */
 867                 struct folio *folio;
 868                 unsigned long addr;
 869
 870                 cond_resched();
 871
 872                 /*
 873                  * fallocate(2) manpage permits EINTR; we may have been
 874                  * interrupted because we are using up too much memory.
 875                  */
 876                 if (signal_pending(current)) {
 877                         error = -EINTR;
 878                         break;
 879                 }
 880
 881                 /* addr is the offset within the file (zero based) */
 882                 addr = index * hpage_size;
 883
 884                 /* mutex taken here, fault path and hole punch */
 885                 hash = hugetlb_fault_mutex_hash(mapping, index);
 886                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 887
 888                 /* See if already present in mapping to avoid alloc/free */
 889                 folio = filemap_get_folio(mapping, index);
 890                 if (!IS_ERR(folio)) {
 891                         folio_put(folio);
 892                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 893                         continue;
 894                 }
 895
 896                 /*
 897                  * Allocate folio without setting the avoid_reserve argument.
 898                  * There certainly are no reserves associated with the
 899                  * pseudo_vma.  However, there could be shared mappings with
 900                  * reserves for the file at the inode level.  If we fallocate
 901                  * folios in these areas, we need to consume the reserves
 902                  * to keep reservation accounting consistent.
 903                  */
 904                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 905                 folio = alloc_hugetlb_folio(&pseudo_vma, addr, 0);
 906                 hugetlb_drop_vma_policy(&pseudo_vma);
 907                 if (IS_ERR(folio)) {
 908                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 909                         error = PTR_ERR(folio);
 910                         goto out;
 911                 }
 912                 clear_huge_page(&folio->page, addr, pages_per_huge_page(h));
 913                 __folio_mark_uptodate(folio);
 914                 error = hugetlb_add_to_page_cache(folio, mapping, index);
 915                 if (unlikely(error)) {
 916                         restore_reserve_on_error(h, &pseudo_vma, addr, folio);
 917                         folio_put(folio);
 918                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 919                         goto out;
 920                 }
 921
 922                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 923
 924                 folio_set_hugetlb_migratable(folio);
 925                 /*
 926                  * folio_unlock because locked by hugetlb_add_to_page_cache()
 927                  * folio_put() due to reference from alloc_hugetlb_folio()
 928                  */
 929                 folio_unlock(folio);
 930                 folio_put(folio);
 931         }
 932
 933         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 934                 i_size_write(inode, offset + len);
 935         inode_set_ctime_current(inode);
 936 out:
 937         inode_unlock(inode);
 938         return error;
 939 }
 940
 941 static int hugetlbfs_setattr(struct mnt_idmap *idmap,
 942                              struct dentry *dentry, struct iattr *attr)
 943 {
 944         struct inode *inode = d_inode(dentry);
 945         struct hstate *h = hstate_inode(inode);
 946         int error;
 947         unsigned int ia_valid = attr->ia_valid;
 948         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 949
 950         error = setattr_prepare(&nop_mnt_idmap, dentry, attr);
 951         if (error)
 952                 return error;
 953
 954         if (ia_valid & ATTR_SIZE) {
 955                 loff_t oldsize = inode->i_size;
 956                 loff_t newsize = attr->ia_size;
 957
 958                 if (newsize & ~huge_page_mask(h))
 959                         return -EINVAL;
 960                 /* protected by i_rwsem */
 961                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 962                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 963                         return -EPERM;
 964                 hugetlb_vmtruncate(inode, newsize);
 965         }
 966
 967         setattr_copy(&nop_mnt_idmap, inode, attr);
 968         mark_inode_dirty(inode);
 969         return 0;
 970 }
 971
 972 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 973                                         struct hugetlbfs_fs_context *ctx)
 974 {
 975         struct inode *inode;
 976
 977         inode = new_inode(sb);
 978         if (inode) {
 979                 inode->i_ino = get_next_ino();
 980                 inode->i_mode = S_IFDIR | ctx->mode;
 981                 inode->i_uid = ctx->uid;
 982                 inode->i_gid = ctx->gid;
 983                 inode->i_atime = inode->i_mtime = inode_set_ctime_current(inode);
 984                 inode->i_op = &hugetlbfs_dir_inode_operations;
 985                 inode->i_fop = &simple_dir_operations;
 986                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 987                 inc_nlink(inode);
 988                 lockdep_annotate_inode_mutex_key(inode);
 989         }
 990         return inode;
 991 }
 992
 993 /*
 994  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 995  * be taken from reclaim -- unlike regular filesystems. This needs an
 996  * annotation because huge_pmd_share() does an allocation under hugetlb's
 997  * i_mmap_rwsem.
 998  */
 999 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
1000
1001 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
1002                                         struct inode *dir,
1003                                         umode_t mode, dev_t dev)
1004 {
1005         struct inode *inode;
1006         struct resv_map *resv_map = NULL;
1007
1008         /*
1009          * Reserve maps are only needed for inodes that can have associated
1010          * page allocations.
1011          */
1012         if (S_ISREG(mode) || S_ISLNK(mode)) {
1013                 resv_map = resv_map_alloc();
1014                 if (!resv_map)
1015                         return NULL;
1016         }
1017
1018         inode = new_inode(sb);
1019         if (inode) {
1020                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
1021
1022                 inode->i_ino = get_next_ino();
1023                 inode_init_owner(&nop_mnt_idmap, inode, dir, mode);
1024                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
1025                                 &hugetlbfs_i_mmap_rwsem_key);
1026                 inode->i_mapping->a_ops = &hugetlbfs_aops;
1027                 inode->i_atime = inode->i_mtime = inode_set_ctime_current(inode);
1028                 inode->i_mapping->private_data = resv_map;
1029                 info->seals = F_SEAL_SEAL;
1030                 switch (mode & S_IFMT) {
1031                 default:
1032                         init_special_inode(inode, mode, dev);
1033                         break;
1034                 case S_IFREG:
1035                         inode->i_op = &hugetlbfs_inode_operations;
1036                         inode->i_fop = &hugetlbfs_file_operations;
1037                         break;
1038                 case S_IFDIR:
1039                         inode->i_op = &hugetlbfs_dir_inode_operations;
1040                         inode->i_fop = &simple_dir_operations;
1041
1042                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
1043                         inc_nlink(inode);
1044                         break;
1045                 case S_IFLNK:
1046                         inode->i_op = &page_symlink_inode_operations;
1047                         inode_nohighmem(inode);
1048                         break;
1049                 }
1050                 lockdep_annotate_inode_mutex_key(inode);
1051         } else {
1052                 if (resv_map)
1053                         kref_put(&resv_map->refs, resv_map_release);
1054         }
1055
1056         return inode;
1057 }
1058
1059 /*
1060  * File creation. Allocate an inode, and we're done..
1061  */
1062 static int hugetlbfs_mknod(struct mnt_idmap *idmap, struct inode *dir,
1063                            struct dentry *dentry, umode_t mode, dev_t dev)
1064 {
1065         struct inode *inode;
1066
1067         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
1068         if (!inode)
1069                 return -ENOSPC;
1070         dir->i_mtime = inode_set_ctime_current(dir);
1071         d_instantiate(dentry, inode);
1072         dget(dentry);/* Extra count - pin the dentry in core */
1073         return 0;
1074 }
1075
1076 static int hugetlbfs_mkdir(struct mnt_idmap *idmap, struct inode *dir,
1077                            struct dentry *dentry, umode_t mode)
1078 {
1079         int retval = hugetlbfs_mknod(&nop_mnt_idmap, dir, dentry,
1080                                      mode | S_IFDIR, 0);
1081         if (!retval)
1082                 inc_nlink(dir);
1083         return retval;
1084 }
1085
1086 static int hugetlbfs_create(struct mnt_idmap *idmap,
1087                             struct inode *dir, struct dentry *dentry,
1088                             umode_t mode, bool excl)
1089 {
1090         return hugetlbfs_mknod(&nop_mnt_idmap, dir, dentry, mode | S_IFREG, 0);
1091 }
1092
1093 static int hugetlbfs_tmpfile(struct mnt_idmap *idmap,
1094                              struct inode *dir, struct file *file,
1095                              umode_t mode)
1096 {
1097         struct inode *inode;
1098
1099         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode | S_IFREG, 0);
1100         if (!inode)
1101                 return -ENOSPC;
1102         dir->i_mtime = inode_set_ctime_current(dir);
1103         d_tmpfile(file, inode);
1104         return finish_open_simple(file, 0);
1105 }
1106
1107 static int hugetlbfs_symlink(struct mnt_idmap *idmap,
1108                              struct inode *dir, struct dentry *dentry,
1109                              const char *symname)
1110 {
1111         struct inode *inode;
1112         int error = -ENOSPC;
1113
1114         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
1115         if (inode) {
1116                 int l = strlen(symname)+1;
1117                 error = page_symlink(inode, symname, l);
1118                 if (!error) {
1119                         d_instantiate(dentry, inode);
1120                         dget(dentry);
1121                 } else
1122                         iput(inode);
1123         }
1124         dir->i_mtime = inode_set_ctime_current(dir);
1125
1126         return error;
1127 }
1128
1129 #ifdef CONFIG_MIGRATION
1130 static int hugetlbfs_migrate_folio(struct address_space *mapping,
1131                                 struct folio *dst, struct folio *src,
1132                                 enum migrate_mode mode)
1133 {
1134         int rc;
1135
1136         rc = migrate_huge_page_move_mapping(mapping, dst, src);
1137         if (rc != MIGRATEPAGE_SUCCESS)
1138                 return rc;
1139
1140         if (hugetlb_folio_subpool(src)) {
1141                 hugetlb_set_folio_subpool(dst,
1142                                         hugetlb_folio_subpool(src));
1143                 hugetlb_set_folio_subpool(src, NULL);
1144         }
1145
1146         if (mode != MIGRATE_SYNC_NO_COPY)
1147                 folio_migrate_copy(dst, src);
1148         else
1149                 folio_migrate_flags(dst, src);
1150
1151         return MIGRATEPAGE_SUCCESS;
1152 }
1153 #else
1154 #define hugetlbfs_migrate_folio NULL
1155 #endif
1156
1157 static int hugetlbfs_error_remove_page(struct address_space *mapping,
1158                                 struct page *page)
1159 {
1160         return 0;
1161 }
1162
1163 /*
1164  * Display the mount options in /proc/mounts.
1165  */
1166 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1167 {
1168         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1169         struct hugepage_subpool *spool = sbinfo->spool;
1170         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1171         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1172         char mod;
1173
1174         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1175                 seq_printf(m, ",uid=%u",
1176                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1177         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1178                 seq_printf(m, ",gid=%u",
1179                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1180         if (sbinfo->mode != 0755)
1181                 seq_printf(m, ",mode=%o", sbinfo->mode);
1182         if (sbinfo->max_inodes != -1)
1183                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1184
1185         hpage_size /= 1024;
1186         mod = 'K';
1187         if (hpage_size >= 1024) {
1188                 hpage_size /= 1024;
1189                 mod = 'M';
1190         }
1191         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1192         if (spool) {
1193                 if (spool->max_hpages != -1)
1194                         seq_printf(m, ",size=%llu",
1195                                    (unsigned long long)spool->max_hpages << hpage_shift);
1196                 if (spool->min_hpages != -1)
1197                         seq_printf(m, ",min_size=%llu",
1198                                    (unsigned long long)spool->min_hpages << hpage_shift);
1199         }
1200         return 0;
1201 }
1202
1203 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1204 {
1205         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1206         struct hstate *h = hstate_inode(d_inode(dentry));
1207
1208         buf->f_type = HUGETLBFS_MAGIC;
1209         buf->f_bsize = huge_page_size(h);
1210         if (sbinfo) {
1211                 spin_lock(&sbinfo->stat_lock);
1212                 /* If no limits set, just report 0 or -1 for max/free/used
1213                  * blocks, like simple_statfs() */
1214                 if (sbinfo->spool) {
1215                         long free_pages;
1216
1217                         spin_lock_irq(&sbinfo->spool->lock);
1218                         buf->f_blocks = sbinfo->spool->max_hpages;
1219                         free_pages = sbinfo->spool->max_hpages
1220                                 - sbinfo->spool->used_hpages;
1221                         buf->f_bavail = buf->f_bfree = free_pages;
1222                         spin_unlock_irq(&sbinfo->spool->lock);
1223                         buf->f_files = sbinfo->max_inodes;
1224                         buf->f_ffree = sbinfo->free_inodes;
1225                 }
1226                 spin_unlock(&sbinfo->stat_lock);
1227         }
1228         buf->f_namelen = NAME_MAX;
1229         return 0;
1230 }
1231
1232 static void hugetlbfs_put_super(struct super_block *sb)
1233 {
1234         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1235
1236         if (sbi) {
1237                 sb->s_fs_info = NULL;
1238
1239                 if (sbi->spool)
1240                         hugepage_put_subpool(sbi->spool);
1241
1242                 kfree(sbi);
1243         }
1244 }
1245
1246 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1247 {
1248         if (sbinfo->free_inodes >= 0) {
1249                 spin_lock(&sbinfo->stat_lock);
1250                 if (unlikely(!sbinfo->free_inodes)) {
1251                         spin_unlock(&sbinfo->stat_lock);
1252                         return 0;
1253                 }
1254                 sbinfo->free_inodes--;
1255                 spin_unlock(&sbinfo->stat_lock);
1256         }
1257
1258         return 1;
1259 }
1260
1261 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1262 {
1263         if (sbinfo->free_inodes >= 0) {
1264                 spin_lock(&sbinfo->stat_lock);
1265                 sbinfo->free_inodes++;
1266                 spin_unlock(&sbinfo->stat_lock);
1267         }
1268 }
1269
1270
1271 static struct kmem_cache *hugetlbfs_inode_cachep;
1272
1273 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1274 {
1275         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1276         struct hugetlbfs_inode_info *p;
1277
1278         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1279                 return NULL;
1280         p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1281         if (unlikely(!p)) {
1282                 hugetlbfs_inc_free_inodes(sbinfo);
1283                 return NULL;
1284         }
1285
1286         /*
1287          * Any time after allocation, hugetlbfs_destroy_inode can be called
1288          * for the inode.  mpol_free_shared_policy is unconditionally called
1289          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1290          * in case of a quick call to destroy.
1291          *
1292          * Note that the policy is initialized even if we are creating a
1293          * private inode.  This simplifies hugetlbfs_destroy_inode.
1294          */
1295         mpol_shared_policy_init(&p->policy, NULL);
1296
1297         return &p->vfs_inode;
1298 }
1299
1300 static void hugetlbfs_free_inode(struct inode *inode)
1301 {
1302         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1303 }
1304
1305 static void hugetlbfs_destroy_inode(struct inode *inode)
1306 {
1307         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1308         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1309 }
1310
1311 static const struct address_space_operations hugetlbfs_aops = {
1312         .write_begin    = hugetlbfs_write_begin,
1313         .write_end      = hugetlbfs_write_end,
1314         .dirty_folio    = noop_dirty_folio,
1315         .migrate_folio  = hugetlbfs_migrate_folio,
1316         .error_remove_page      = hugetlbfs_error_remove_page,
1317 };
1318
1319
1320 static void init_once(void *foo)
1321 {
1322         struct hugetlbfs_inode_info *ei = foo;
1323
1324         inode_init_once(&ei->vfs_inode);
1325 }
1326
1327 const struct file_operations hugetlbfs_file_operations = {
1328         .read_iter              = hugetlbfs_read_iter,
1329         .mmap                   = hugetlbfs_file_mmap,
1330         .fsync                  = noop_fsync,
1331         .get_unmapped_area      = hugetlb_get_unmapped_area,
1332         .llseek                 = default_llseek,
1333         .fallocate              = hugetlbfs_fallocate,
1334 };
1335
1336 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1337         .create         = hugetlbfs_create,
1338         .lookup         = simple_lookup,
1339         .link           = simple_link,
1340         .unlink         = simple_unlink,
1341         .symlink        = hugetlbfs_symlink,
1342         .mkdir          = hugetlbfs_mkdir,
1343         .rmdir          = simple_rmdir,
1344         .mknod          = hugetlbfs_mknod,
1345         .rename         = simple_rename,
1346         .setattr        = hugetlbfs_setattr,
1347         .tmpfile        = hugetlbfs_tmpfile,
1348 };
1349
1350 static const struct inode_operations hugetlbfs_inode_operations = {
1351         .setattr        = hugetlbfs_setattr,
1352 };
1353
1354 static const struct super_operations hugetlbfs_ops = {
1355         .alloc_inode    = hugetlbfs_alloc_inode,
1356         .free_inode     = hugetlbfs_free_inode,
1357         .destroy_inode  = hugetlbfs_destroy_inode,
1358         .evict_inode    = hugetlbfs_evict_inode,
1359         .statfs         = hugetlbfs_statfs,
1360         .put_super      = hugetlbfs_put_super,
1361         .show_options   = hugetlbfs_show_options,
1362 };
1363
1364 /*
1365  * Convert size option passed from command line to number of huge pages
1366  * in the pool specified by hstate.  Size option could be in bytes
1367  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1368  */
1369 static long
1370 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1371                          enum hugetlbfs_size_type val_type)
1372 {
1373         if (val_type == NO_SIZE)
1374                 return -1;
1375
1376         if (val_type == SIZE_PERCENT) {
1377                 size_opt <<= huge_page_shift(h);
1378                 size_opt *= h->max_huge_pages;
1379                 do_div(size_opt, 100);
1380         }
1381
1382         size_opt >>= huge_page_shift(h);
1383         return size_opt;
1384 }
1385
1386 /*
1387  * Parse one mount parameter.
1388  */
1389 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1390 {
1391         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1392         struct fs_parse_result result;
1393         char *rest;
1394         unsigned long ps;
1395         int opt;
1396
1397         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1398         if (opt < 0)
1399                 return opt;
1400
1401         switch (opt) {
1402         case Opt_uid:
1403                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1404                 if (!uid_valid(ctx->uid))
1405                         goto bad_val;
1406                 return 0;
1407
1408         case Opt_gid:
1409                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1410                 if (!gid_valid(ctx->gid))
1411                         goto bad_val;
1412                 return 0;
1413
1414         case Opt_mode:
1415                 ctx->mode = result.uint_32 & 01777U;
1416                 return 0;
1417
1418         case Opt_size:
1419                 /* memparse() will accept a K/M/G without a digit */
1420                 if (!param->string || !isdigit(param->string[0]))
1421                         goto bad_val;
1422                 ctx->max_size_opt = memparse(param->string, &rest);
1423                 ctx->max_val_type = SIZE_STD;
1424                 if (*rest == '%')
1425                         ctx->max_val_type = SIZE_PERCENT;
1426                 return 0;
1427
1428         case Opt_nr_inodes:
1429                 /* memparse() will accept a K/M/G without a digit */
1430                 if (!param->string || !isdigit(param->string[0]))
1431                         goto bad_val;
1432                 ctx->nr_inodes = memparse(param->string, &rest);
1433                 return 0;
1434
1435         case Opt_pagesize:
1436                 ps = memparse(param->string, &rest);
1437                 ctx->hstate = size_to_hstate(ps);
1438                 if (!ctx->hstate) {
1439                         pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1440                         return -EINVAL;
1441                 }
1442                 return 0;
1443
1444         case Opt_min_size:
1445                 /* memparse() will accept a K/M/G without a digit */
1446                 if (!param->string || !isdigit(param->string[0]))
1447                         goto bad_val;
1448                 ctx->min_size_opt = memparse(param->string, &rest);
1449                 ctx->min_val_type = SIZE_STD;
1450                 if (*rest == '%')
1451                         ctx->min_val_type = SIZE_PERCENT;
1452                 return 0;
1453
1454         default:
1455                 return -EINVAL;
1456         }
1457
1458 bad_val:
1459         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1460                       param->string, param->key);
1461 }
1462
1463 /*
1464  * Validate the parsed options.
1465  */
1466 static int hugetlbfs_validate(struct fs_context *fc)
1467 {
1468         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1469
1470         /*
1471          * Use huge page pool size (in hstate) to convert the size
1472          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1473          */
1474         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1475                                                    ctx->max_size_opt,
1476                                                    ctx->max_val_type);
1477         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1478                                                    ctx->min_size_opt,
1479                                                    ctx->min_val_type);
1480
1481         /*
1482          * If max_size was specified, then min_size must be smaller
1483          */
1484         if (ctx->max_val_type > NO_SIZE &&
1485             ctx->min_hpages > ctx->max_hpages) {
1486                 pr_err("Minimum size can not be greater than maximum size\n");
1487                 return -EINVAL;
1488         }
1489
1490         return 0;
1491 }
1492
1493 static int
1494 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1495 {
1496         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1497         struct hugetlbfs_sb_info *sbinfo;
1498
1499         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1500         if (!sbinfo)
1501                 return -ENOMEM;
1502         sb->s_fs_info = sbinfo;
1503         spin_lock_init(&sbinfo->stat_lock);
1504         sbinfo->hstate          = ctx->hstate;
1505         sbinfo->max_inodes      = ctx->nr_inodes;
1506         sbinfo->free_inodes     = ctx->nr_inodes;
1507         sbinfo->spool           = NULL;
1508         sbinfo->uid             = ctx->uid;
1509         sbinfo->gid             = ctx->gid;
1510         sbinfo->mode            = ctx->mode;
1511
1512         /*
1513          * Allocate and initialize subpool if maximum or minimum size is
1514          * specified.  Any needed reservations (for minimum size) are taken
1515          * when the subpool is created.
1516          */
1517         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1518                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1519                                                      ctx->max_hpages,
1520                                                      ctx->min_hpages);
1521                 if (!sbinfo->spool)
1522                         goto out_free;
1523         }
1524         sb->s_maxbytes = MAX_LFS_FILESIZE;
1525         sb->s_blocksize = huge_page_size(ctx->hstate);
1526         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1527         sb->s_magic = HUGETLBFS_MAGIC;
1528         sb->s_op = &hugetlbfs_ops;
1529         sb->s_time_gran = 1;
1530
1531         /*
1532          * Due to the special and limited functionality of hugetlbfs, it does
1533          * not work well as a stacking filesystem.
1534          */
1535         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1536         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1537         if (!sb->s_root)
1538                 goto out_free;
1539         return 0;
1540 out_free:
1541         kfree(sbinfo->spool);
1542         kfree(sbinfo);
1543         return -ENOMEM;
1544 }
1545
1546 static int hugetlbfs_get_tree(struct fs_context *fc)
1547 {
1548         int err = hugetlbfs_validate(fc);
1549         if (err)
1550                 return err;
1551         return get_tree_nodev(fc, hugetlbfs_fill_super);
1552 }
1553
1554 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1555 {
1556         kfree(fc->fs_private);
1557 }
1558
1559 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1560         .free           = hugetlbfs_fs_context_free,
1561         .parse_param    = hugetlbfs_parse_param,
1562         .get_tree       = hugetlbfs_get_tree,
1563 };
1564
1565 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1566 {
1567         struct hugetlbfs_fs_context *ctx;
1568
1569         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1570         if (!ctx)
1571                 return -ENOMEM;
1572
1573         ctx->max_hpages = -1; /* No limit on size by default */
1574         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1575         ctx->uid        = current_fsuid();
1576         ctx->gid        = current_fsgid();
1577         ctx->mode       = 0755;
1578         ctx->hstate     = &default_hstate;
1579         ctx->min_hpages = -1; /* No default minimum size */
1580         ctx->max_val_type = NO_SIZE;
1581         ctx->min_val_type = NO_SIZE;
1582         fc->fs_private = ctx;
1583         fc->ops = &hugetlbfs_fs_context_ops;
1584         return 0;
1585 }
1586
1587 static struct file_system_type hugetlbfs_fs_type = {
1588         .name                   = "hugetlbfs",
1589         .init_fs_context        = hugetlbfs_init_fs_context,
1590         .parameters             = hugetlb_fs_parameters,
1591         .kill_sb                = kill_litter_super,
1592 };
1593
1594 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1595
1596 static int can_do_hugetlb_shm(void)
1597 {
1598         kgid_t shm_group;
1599         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1600         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1601 }
1602
1603 static int get_hstate_idx(int page_size_log)
1604 {
1605         struct hstate *h = hstate_sizelog(page_size_log);
1606
1607         if (!h)
1608                 return -1;
1609         return hstate_index(h);
1610 }
1611
1612 /*
1613  * Note that size should be aligned to proper hugepage size in caller side,
1614  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1615  */
1616 struct file *hugetlb_file_setup(const char *name, size_t size,
1617                                 vm_flags_t acctflag, int creat_flags,
1618                                 int page_size_log)
1619 {
1620         struct inode *inode;
1621         struct vfsmount *mnt;
1622         int hstate_idx;
1623         struct file *file;
1624
1625         hstate_idx = get_hstate_idx(page_size_log);
1626         if (hstate_idx < 0)
1627                 return ERR_PTR(-ENODEV);
1628
1629         mnt = hugetlbfs_vfsmount[hstate_idx];
1630         if (!mnt)
1631                 return ERR_PTR(-ENOENT);
1632
1633         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1634                 struct ucounts *ucounts = current_ucounts();
1635
1636                 if (user_shm_lock(size, ucounts)) {
1637                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1638                                 current->comm, current->pid);
1639                         user_shm_unlock(size, ucounts);
1640                 }
1641                 return ERR_PTR(-EPERM);
1642         }
1643
1644         file = ERR_PTR(-ENOSPC);
1645         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1646         if (!inode)
1647                 goto out;
1648         if (creat_flags == HUGETLB_SHMFS_INODE)
1649                 inode->i_flags |= S_PRIVATE;
1650
1651         inode->i_size = size;
1652         clear_nlink(inode);
1653
1654         if (!hugetlb_reserve_pages(inode, 0,
1655                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1656                         acctflag))
1657                 file = ERR_PTR(-ENOMEM);
1658         else
1659                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1660                                         &hugetlbfs_file_operations);
1661         if (!IS_ERR(file))
1662                 return file;
1663
1664         iput(inode);
1665 out:
1666         return file;
1667 }
1668
1669 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1670 {
1671         struct fs_context *fc;
1672         struct vfsmount *mnt;
1673
1674         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1675         if (IS_ERR(fc)) {
1676                 mnt = ERR_CAST(fc);
1677         } else {
1678                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1679                 ctx->hstate = h;
1680                 mnt = fc_mount(fc);
1681                 put_fs_context(fc);
1682         }
1683         if (IS_ERR(mnt))
1684                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1685                        huge_page_size(h) / SZ_1K);
1686         return mnt;
1687 }
1688
1689 static int __init init_hugetlbfs_fs(void)
1690 {
1691         struct vfsmount *mnt;
1692         struct hstate *h;
1693         int error;
1694         int i;
1695
1696         if (!hugepages_supported()) {
1697                 pr_info("disabling because there are no supported hugepage sizes\n");
1698                 return -ENOTSUPP;
1699         }
1700
1701         error = -ENOMEM;
1702         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1703                                         sizeof(struct hugetlbfs_inode_info),
1704                                         0, SLAB_ACCOUNT, init_once);
1705         if (hugetlbfs_inode_cachep == NULL)
1706                 goto out;
1707
1708         error = register_filesystem(&hugetlbfs_fs_type);
1709         if (error)
1710                 goto out_free;
1711
1712         /* default hstate mount is required */
1713         mnt = mount_one_hugetlbfs(&default_hstate);
1714         if (IS_ERR(mnt)) {
1715                 error = PTR_ERR(mnt);
1716                 goto out_unreg;
1717         }
1718         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1719
1720         /* other hstates are optional */
1721         i = 0;
1722         for_each_hstate(h) {
1723                 if (i == default_hstate_idx) {
1724                         i++;
1725                         continue;
1726                 }
1727
1728                 mnt = mount_one_hugetlbfs(h);
1729                 if (IS_ERR(mnt))
1730                         hugetlbfs_vfsmount[i] = NULL;
1731                 else
1732                         hugetlbfs_vfsmount[i] = mnt;
1733                 i++;
1734         }
1735
1736         return 0;
1737
1738  out_unreg:
1739         (void)unregister_filesystem(&hugetlbfs_fs_type);
1740  out_free:
1741         kmem_cache_destroy(hugetlbfs_inode_cachep);
1742  out:
1743         return error;
1744 }
1745 fs_initcall(init_hugetlbfs_fs)