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