fs/ext4/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  linux/fs/ext4/file.c
   4  *
   5  * Copyright (C) 1992, 1993, 1994, 1995
   6  * Remy Card ([email protected])
   7  * Laboratoire MASI - Institut Blaise Pascal
   8  * Universite Pierre et Marie Curie (Paris VI)
   9  *
  10  *  from
  11  *
  12  *  linux/fs/minix/file.c
  13  *
  14  *  Copyright (C) 1991, 1992  Linus Torvalds
  15  *
  16  *  ext4 fs regular file handling primitives
  17  *
  18  *  64-bit file support on 64-bit platforms by Jakub Jelinek
  19  *      ([email protected])
  20  */
  21
  22 #include <linux/time.h>
  23 #include <linux/fs.h>
  24 #include <linux/iomap.h>
  25 #include <linux/mount.h>
  26 #include <linux/path.h>
  27 #include <linux/dax.h>
  28 #include <linux/quotaops.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/uio.h>
  31 #include <linux/mman.h>
  32 #include <linux/backing-dev.h>
  33 #include "ext4.h"
  34 #include "ext4_jbd2.h"
  35 #include "xattr.h"
  36 #include "acl.h"
  37 #include "truncate.h"
  38
  39 /*
  40  * Returns %true if the given DIO request should be attempted with DIO, or
  41  * %false if it should fall back to buffered I/O.
  42  *
  43  * DIO isn't well specified; when it's unsupported (either due to the request
  44  * being misaligned, or due to the file not supporting DIO at all), filesystems
  45  * either fall back to buffered I/O or return EINVAL.  For files that don't use
  46  * any special features like encryption or verity, ext4 has traditionally
  47  * returned EINVAL for misaligned DIO.  iomap_dio_rw() uses this convention too.
  48  * In this case, we should attempt the DIO, *not* fall back to buffered I/O.
  49  *
  50  * In contrast, in cases where DIO is unsupported due to ext4 features, ext4
  51  * traditionally falls back to buffered I/O.
  52  *
  53  * This function implements the traditional ext4 behavior in all these cases.
  54  */
  55 static bool ext4_should_use_dio(struct kiocb *iocb, struct iov_iter *iter)
  56 {
  57         struct inode *inode = file_inode(iocb->ki_filp);
  58         u32 dio_align = ext4_dio_alignment(inode);
  59
  60         if (dio_align == 0)
  61                 return false;
  62
  63         if (dio_align == 1)
  64                 return true;
  65
  66         return IS_ALIGNED(iocb->ki_pos | iov_iter_alignment(iter), dio_align);
  67 }
  68
  69 static ssize_t ext4_dio_read_iter(struct kiocb *iocb, struct iov_iter *to)
  70 {
  71         ssize_t ret;
  72         struct inode *inode = file_inode(iocb->ki_filp);
  73
  74         if (iocb->ki_flags & IOCB_NOWAIT) {
  75                 if (!inode_trylock_shared(inode))
  76                         return -EAGAIN;
  77         } else {
  78                 inode_lock_shared(inode);
  79         }
  80
  81         if (!ext4_should_use_dio(iocb, to)) {
  82                 inode_unlock_shared(inode);
  83                 /*
  84                  * Fallback to buffered I/O if the operation being performed on
  85                  * the inode is not supported by direct I/O. The IOCB_DIRECT
  86                  * flag needs to be cleared here in order to ensure that the
  87                  * direct I/O path within generic_file_read_iter() is not
  88                  * taken.
  89                  */
  90                 iocb->ki_flags &= ~IOCB_DIRECT;
  91                 return generic_file_read_iter(iocb, to);
  92         }
  93
  94         ret = iomap_dio_rw(iocb, to, &ext4_iomap_ops, NULL, 0, NULL, 0);
  95         inode_unlock_shared(inode);
  96
  97         file_accessed(iocb->ki_filp);
  98         return ret;
  99 }
 100
 101 #ifdef CONFIG_FS_DAX
 102 static ssize_t ext4_dax_read_iter(struct kiocb *iocb, struct iov_iter *to)
 103 {
 104         struct inode *inode = file_inode(iocb->ki_filp);
 105         ssize_t ret;
 106
 107         if (iocb->ki_flags & IOCB_NOWAIT) {
 108                 if (!inode_trylock_shared(inode))
 109                         return -EAGAIN;
 110         } else {
 111                 inode_lock_shared(inode);
 112         }
 113         /*
 114          * Recheck under inode lock - at this point we are sure it cannot
 115          * change anymore
 116          */
 117         if (!IS_DAX(inode)) {
 118                 inode_unlock_shared(inode);
 119                 /* Fallback to buffered IO in case we cannot support DAX */
 120                 return generic_file_read_iter(iocb, to);
 121         }
 122         ret = dax_iomap_rw(iocb, to, &ext4_iomap_ops);
 123         inode_unlock_shared(inode);
 124
 125         file_accessed(iocb->ki_filp);
 126         return ret;
 127 }
 128 #endif
 129
 130 static ssize_t ext4_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 131 {
 132         struct inode *inode = file_inode(iocb->ki_filp);
 133
 134         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 135                 return -EIO;
 136
 137         if (!iov_iter_count(to))
 138                 return 0; /* skip atime */
 139
 140 #ifdef CONFIG_FS_DAX
 141         if (IS_DAX(inode))
 142                 return ext4_dax_read_iter(iocb, to);
 143 #endif
 144         if (iocb->ki_flags & IOCB_DIRECT)
 145                 return ext4_dio_read_iter(iocb, to);
 146
 147         return generic_file_read_iter(iocb, to);
 148 }
 149
 150 /*
 151  * Called when an inode is released. Note that this is different
 152  * from ext4_file_open: open gets called at every open, but release
 153  * gets called only when /all/ the files are closed.
 154  */
 155 static int ext4_release_file(struct inode *inode, struct file *filp)
 156 {
 157         if (ext4_test_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE)) {
 158                 ext4_alloc_da_blocks(inode);
 159                 ext4_clear_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
 160         }
 161         /* if we are the last writer on the inode, drop the block reservation */
 162         if ((filp->f_mode & FMODE_WRITE) &&
 163                         (atomic_read(&inode->i_writecount) == 1) &&
 164                         !EXT4_I(inode)->i_reserved_data_blocks) {
 165                 down_write(&EXT4_I(inode)->i_data_sem);
 166                 ext4_discard_preallocations(inode, 0);
 167                 up_write(&EXT4_I(inode)->i_data_sem);
 168         }
 169         if (is_dx(inode) && filp->private_data)
 170                 ext4_htree_free_dir_info(filp->private_data);
 171
 172         return 0;
 173 }
 174
 175 /*
 176  * This tests whether the IO in question is block-aligned or not.
 177  * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
 178  * are converted to written only after the IO is complete.  Until they are
 179  * mapped, these blocks appear as holes, so dio_zero_block() will assume that
 180  * it needs to zero out portions of the start and/or end block.  If 2 AIO
 181  * threads are at work on the same unwritten block, they must be synchronized
 182  * or one thread will zero the other's data, causing corruption.
 183  */
 184 static bool
 185 ext4_unaligned_io(struct inode *inode, struct iov_iter *from, loff_t pos)
 186 {
 187         struct super_block *sb = inode->i_sb;
 188         unsigned long blockmask = sb->s_blocksize - 1;
 189
 190         if ((pos | iov_iter_alignment(from)) & blockmask)
 191                 return true;
 192
 193         return false;
 194 }
 195
 196 static bool
 197 ext4_extending_io(struct inode *inode, loff_t offset, size_t len)
 198 {
 199         if (offset + len > i_size_read(inode) ||
 200             offset + len > EXT4_I(inode)->i_disksize)
 201                 return true;
 202         return false;
 203 }
 204
 205 /* Is IO overwriting allocated or initialized blocks? */
 206 static bool ext4_overwrite_io(struct inode *inode,
 207                               loff_t pos, loff_t len, bool *unwritten)
 208 {
 209         struct ext4_map_blocks map;
 210         unsigned int blkbits = inode->i_blkbits;
 211         int err, blklen;
 212
 213         if (pos + len > i_size_read(inode))
 214                 return false;
 215
 216         map.m_lblk = pos >> blkbits;
 217         map.m_len = EXT4_MAX_BLOCKS(len, pos, blkbits);
 218         blklen = map.m_len;
 219
 220         err = ext4_map_blocks(NULL, inode, &map, 0);
 221         if (err != blklen)
 222                 return false;
 223         /*
 224          * 'err==len' means that all of the blocks have been preallocated,
 225          * regardless of whether they have been initialized or not. We need to
 226          * check m_flags to distinguish the unwritten extents.
 227          */
 228         *unwritten = !(map.m_flags & EXT4_MAP_MAPPED);
 229         return true;
 230 }
 231
 232 static ssize_t ext4_generic_write_checks(struct kiocb *iocb,
 233                                          struct iov_iter *from)
 234 {
 235         struct inode *inode = file_inode(iocb->ki_filp);
 236         ssize_t ret;
 237
 238         if (unlikely(IS_IMMUTABLE(inode)))
 239                 return -EPERM;
 240
 241         ret = generic_write_checks(iocb, from);
 242         if (ret <= 0)
 243                 return ret;
 244
 245         /*
 246          * If we have encountered a bitmap-format file, the size limit
 247          * is smaller than s_maxbytes, which is for extent-mapped files.
 248          */
 249         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
 250                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 251
 252                 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes)
 253                         return -EFBIG;
 254                 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
 255         }
 256
 257         return iov_iter_count(from);
 258 }
 259
 260 static ssize_t ext4_write_checks(struct kiocb *iocb, struct iov_iter *from)
 261 {
 262         ssize_t ret, count;
 263
 264         count = ext4_generic_write_checks(iocb, from);
 265         if (count <= 0)
 266                 return count;
 267
 268         ret = file_modified(iocb->ki_filp);
 269         if (ret)
 270                 return ret;
 271         return count;
 272 }
 273
 274 static ssize_t ext4_buffered_write_iter(struct kiocb *iocb,
 275                                         struct iov_iter *from)
 276 {
 277         ssize_t ret;
 278         struct inode *inode = file_inode(iocb->ki_filp);
 279
 280         if (iocb->ki_flags & IOCB_NOWAIT)
 281                 return -EOPNOTSUPP;
 282
 283         inode_lock(inode);
 284         ret = ext4_write_checks(iocb, from);
 285         if (ret <= 0)
 286                 goto out;
 287
 288         current->backing_dev_info = inode_to_bdi(inode);
 289         ret = generic_perform_write(iocb, from);
 290         current->backing_dev_info = NULL;
 291
 292 out:
 293         inode_unlock(inode);
 294         if (likely(ret > 0)) {
 295                 iocb->ki_pos += ret;
 296                 ret = generic_write_sync(iocb, ret);
 297         }
 298
 299         return ret;
 300 }
 301
 302 static ssize_t ext4_handle_inode_extension(struct inode *inode, loff_t offset,
 303                                            ssize_t written, size_t count)
 304 {
 305         handle_t *handle;
 306         bool truncate = false;
 307         u8 blkbits = inode->i_blkbits;
 308         ext4_lblk_t written_blk, end_blk;
 309         int ret;
 310
 311         /*
 312          * Note that EXT4_I(inode)->i_disksize can get extended up to
 313          * inode->i_size while the I/O was running due to writeback of delalloc
 314          * blocks. But, the code in ext4_iomap_alloc() is careful to use
 315          * zeroed/unwritten extents if this is possible; thus we won't leave
 316          * uninitialized blocks in a file even if we didn't succeed in writing
 317          * as much as we intended.
 318          */
 319         WARN_ON_ONCE(i_size_read(inode) < EXT4_I(inode)->i_disksize);
 320         if (offset + count <= EXT4_I(inode)->i_disksize) {
 321                 /*
 322                  * We need to ensure that the inode is removed from the orphan
 323                  * list if it has been added prematurely, due to writeback of
 324                  * delalloc blocks.
 325                  */
 326                 if (!list_empty(&EXT4_I(inode)->i_orphan) && inode->i_nlink) {
 327                         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 328
 329                         if (IS_ERR(handle)) {
 330                                 ext4_orphan_del(NULL, inode);
 331                                 return PTR_ERR(handle);
 332                         }
 333
 334                         ext4_orphan_del(handle, inode);
 335                         ext4_journal_stop(handle);
 336                 }
 337
 338                 return written;
 339         }
 340
 341         if (written < 0)
 342                 goto truncate;
 343
 344         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 345         if (IS_ERR(handle)) {
 346                 written = PTR_ERR(handle);
 347                 goto truncate;
 348         }
 349
 350         if (ext4_update_inode_size(inode, offset + written)) {
 351                 ret = ext4_mark_inode_dirty(handle, inode);
 352                 if (unlikely(ret)) {
 353                         written = ret;
 354                         ext4_journal_stop(handle);
 355                         goto truncate;
 356                 }
 357         }
 358
 359         /*
 360          * We may need to truncate allocated but not written blocks beyond EOF.
 361          */
 362         written_blk = ALIGN(offset + written, 1 << blkbits);
 363         end_blk = ALIGN(offset + count, 1 << blkbits);
 364         if (written_blk < end_blk && ext4_can_truncate(inode))
 365                 truncate = true;
 366
 367         /*
 368          * Remove the inode from the orphan list if it has been extended and
 369          * everything went OK.
 370          */
 371         if (!truncate && inode->i_nlink)
 372                 ext4_orphan_del(handle, inode);
 373         ext4_journal_stop(handle);
 374
 375         if (truncate) {
 376 truncate:
 377                 ext4_truncate_failed_write(inode);
 378                 /*
 379                  * If the truncate operation failed early, then the inode may
 380                  * still be on the orphan list. In that case, we need to try
 381                  * remove the inode from the in-memory linked list.
 382                  */
 383                 if (inode->i_nlink)
 384                         ext4_orphan_del(NULL, inode);
 385         }
 386
 387         return written;
 388 }
 389
 390 static int ext4_dio_write_end_io(struct kiocb *iocb, ssize_t size,
 391                                  int error, unsigned int flags)
 392 {
 393         loff_t pos = iocb->ki_pos;
 394         struct inode *inode = file_inode(iocb->ki_filp);
 395
 396         if (error)
 397                 return error;
 398
 399         if (size && flags & IOMAP_DIO_UNWRITTEN) {
 400                 error = ext4_convert_unwritten_extents(NULL, inode, pos, size);
 401                 if (error < 0)
 402                         return error;
 403         }
 404         /*
 405          * If we are extending the file, we have to update i_size here before
 406          * page cache gets invalidated in iomap_dio_rw(). Otherwise racing
 407          * buffered reads could zero out too much from page cache pages. Update
 408          * of on-disk size will happen later in ext4_dio_write_iter() where
 409          * we have enough information to also perform orphan list handling etc.
 410          * Note that we perform all extending writes synchronously under
 411          * i_rwsem held exclusively so i_size update is safe here in that case.
 412          * If the write was not extending, we cannot see pos > i_size here
 413          * because operations reducing i_size like truncate wait for all
 414          * outstanding DIO before updating i_size.
 415          */
 416         pos += size;
 417         if (pos > i_size_read(inode))
 418                 i_size_write(inode, pos);
 419
 420         return 0;
 421 }
 422
 423 static const struct iomap_dio_ops ext4_dio_write_ops = {
 424         .end_io = ext4_dio_write_end_io,
 425 };
 426
 427 /*
 428  * The intention here is to start with shared lock acquired then see if any
 429  * condition requires an exclusive inode lock. If yes, then we restart the
 430  * whole operation by releasing the shared lock and acquiring exclusive lock.
 431  *
 432  * - For unaligned_io we never take shared lock as it may cause data corruption
 433  *   when two unaligned IO tries to modify the same block e.g. while zeroing.
 434  *
 435  * - For extending writes case we don't take the shared lock, since it requires
 436  *   updating inode i_disksize and/or orphan handling with exclusive lock.
 437  *
 438  * - shared locking will only be true mostly with overwrites, including
 439  *   initialized blocks and unwritten blocks. For overwrite unwritten blocks
 440  *   we protect splitting extents by i_data_sem in ext4_inode_info, so we can
 441  *   also release exclusive i_rwsem lock.
 442  *
 443  * - Otherwise we will switch to exclusive i_rwsem lock.
 444  */
 445 static ssize_t ext4_dio_write_checks(struct kiocb *iocb, struct iov_iter *from,
 446                                      bool *ilock_shared, bool *extend,
 447                                      bool *unwritten)
 448 {
 449         struct file *file = iocb->ki_filp;
 450         struct inode *inode = file_inode(file);
 451         loff_t offset;
 452         size_t count;
 453         ssize_t ret;
 454
 455 restart:
 456         ret = ext4_generic_write_checks(iocb, from);
 457         if (ret <= 0)
 458                 goto out;
 459
 460         offset = iocb->ki_pos;
 461         count = ret;
 462         if (ext4_extending_io(inode, offset, count))
 463                 *extend = true;
 464         /*
 465          * Determine whether the IO operation will overwrite allocated
 466          * and initialized blocks.
 467          * We need exclusive i_rwsem for changing security info
 468          * in file_modified().
 469          */
 470         if (*ilock_shared && (!IS_NOSEC(inode) || *extend ||
 471              !ext4_overwrite_io(inode, offset, count, unwritten))) {
 472                 if (iocb->ki_flags & IOCB_NOWAIT) {
 473                         ret = -EAGAIN;
 474                         goto out;
 475                 }
 476                 inode_unlock_shared(inode);
 477                 *ilock_shared = false;
 478                 inode_lock(inode);
 479                 goto restart;
 480         }
 481
 482         ret = file_modified(file);
 483         if (ret < 0)
 484                 goto out;
 485
 486         return count;
 487 out:
 488         if (*ilock_shared)
 489                 inode_unlock_shared(inode);
 490         else
 491                 inode_unlock(inode);
 492         return ret;
 493 }
 494
 495 static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 496 {
 497         ssize_t ret;
 498         handle_t *handle;
 499         struct inode *inode = file_inode(iocb->ki_filp);
 500         loff_t offset = iocb->ki_pos;
 501         size_t count = iov_iter_count(from);
 502         const struct iomap_ops *iomap_ops = &ext4_iomap_ops;
 503         bool extend = false, unaligned_io = false, unwritten = false;
 504         bool ilock_shared = true;
 505
 506         /*
 507          * We initially start with shared inode lock unless it is
 508          * unaligned IO which needs exclusive lock anyways.
 509          */
 510         if (ext4_unaligned_io(inode, from, offset)) {
 511                 unaligned_io = true;
 512                 ilock_shared = false;
 513         }
 514         /*
 515          * Quick check here without any i_rwsem lock to see if it is extending
 516          * IO. A more reliable check is done in ext4_dio_write_checks() with
 517          * proper locking in place.
 518          */
 519         if (offset + count > i_size_read(inode))
 520                 ilock_shared = false;
 521
 522         if (iocb->ki_flags & IOCB_NOWAIT) {
 523                 if (ilock_shared) {
 524                         if (!inode_trylock_shared(inode))
 525                                 return -EAGAIN;
 526                 } else {
 527                         if (!inode_trylock(inode))
 528                                 return -EAGAIN;
 529                 }
 530         } else {
 531                 if (ilock_shared)
 532                         inode_lock_shared(inode);
 533                 else
 534                         inode_lock(inode);
 535         }
 536
 537         /* Fallback to buffered I/O if the inode does not support direct I/O. */
 538         if (!ext4_should_use_dio(iocb, from)) {
 539                 if (ilock_shared)
 540                         inode_unlock_shared(inode);
 541                 else
 542                         inode_unlock(inode);
 543                 return ext4_buffered_write_iter(iocb, from);
 544         }
 545
 546         ret = ext4_dio_write_checks(iocb, from,
 547                                     &ilock_shared, &extend, &unwritten);
 548         if (ret <= 0)
 549                 return ret;
 550
 551         /* if we're going to block and IOCB_NOWAIT is set, return -EAGAIN */
 552         if ((iocb->ki_flags & IOCB_NOWAIT) && (unaligned_io || extend)) {
 553                 ret = -EAGAIN;
 554                 goto out;
 555         }
 556         /*
 557          * Make sure inline data cannot be created anymore since we are going
 558          * to allocate blocks for DIO. We know the inode does not have any
 559          * inline data now because ext4_dio_supported() checked for that.
 560          */
 561         ext4_clear_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
 562
 563         offset = iocb->ki_pos;
 564         count = ret;
 565
 566         /*
 567          * Unaligned direct IO must be serialized among each other as zeroing
 568          * of partial blocks of two competing unaligned IOs can result in data
 569          * corruption.
 570          *
 571          * So we make sure we don't allow any unaligned IO in flight.
 572          * For IOs where we need not wait (like unaligned non-AIO DIO),
 573          * below inode_dio_wait() may anyway become a no-op, since we start
 574          * with exclusive lock.
 575          */
 576         if (unaligned_io)
 577                 inode_dio_wait(inode);
 578
 579         if (extend) {
 580                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 581                 if (IS_ERR(handle)) {
 582                         ret = PTR_ERR(handle);
 583                         goto out;
 584                 }
 585
 586                 ret = ext4_orphan_add(handle, inode);
 587                 if (ret) {
 588                         ext4_journal_stop(handle);
 589                         goto out;
 590                 }
 591
 592                 ext4_journal_stop(handle);
 593         }
 594
 595         if (ilock_shared && !unwritten)
 596                 iomap_ops = &ext4_iomap_overwrite_ops;
 597         ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 598                            (unaligned_io || extend) ? IOMAP_DIO_FORCE_WAIT : 0,
 599                            NULL, 0);
 600         if (ret == -ENOTBLK)
 601                 ret = 0;
 602
 603         if (extend)
 604                 ret = ext4_handle_inode_extension(inode, offset, ret, count);
 605
 606 out:
 607         if (ilock_shared)
 608                 inode_unlock_shared(inode);
 609         else
 610                 inode_unlock(inode);
 611
 612         if (ret >= 0 && iov_iter_count(from)) {
 613                 ssize_t err;
 614                 loff_t endbyte;
 615
 616                 offset = iocb->ki_pos;
 617                 err = ext4_buffered_write_iter(iocb, from);
 618                 if (err < 0)
 619                         return err;
 620
 621                 /*
 622                  * We need to ensure that the pages within the page cache for
 623                  * the range covered by this I/O are written to disk and
 624                  * invalidated. This is in attempt to preserve the expected
 625                  * direct I/O semantics in the case we fallback to buffered I/O
 626                  * to complete off the I/O request.
 627                  */
 628                 ret += err;
 629                 endbyte = offset + err - 1;
 630                 err = filemap_write_and_wait_range(iocb->ki_filp->f_mapping,
 631                                                    offset, endbyte);
 632                 if (!err)
 633                         invalidate_mapping_pages(iocb->ki_filp->f_mapping,
 634                                                  offset >> PAGE_SHIFT,
 635                                                  endbyte >> PAGE_SHIFT);
 636         }
 637
 638         return ret;
 639 }
 640
 641 #ifdef CONFIG_FS_DAX
 642 static ssize_t
 643 ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
 644 {
 645         ssize_t ret;
 646         size_t count;
 647         loff_t offset;
 648         handle_t *handle;
 649         bool extend = false;
 650         struct inode *inode = file_inode(iocb->ki_filp);
 651
 652         if (iocb->ki_flags & IOCB_NOWAIT) {
 653                 if (!inode_trylock(inode))
 654                         return -EAGAIN;
 655         } else {
 656                 inode_lock(inode);
 657         }
 658
 659         ret = ext4_write_checks(iocb, from);
 660         if (ret <= 0)
 661                 goto out;
 662
 663         offset = iocb->ki_pos;
 664         count = iov_iter_count(from);
 665
 666         if (offset + count > EXT4_I(inode)->i_disksize) {
 667                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 668                 if (IS_ERR(handle)) {
 669                         ret = PTR_ERR(handle);
 670                         goto out;
 671                 }
 672
 673                 ret = ext4_orphan_add(handle, inode);
 674                 if (ret) {
 675                         ext4_journal_stop(handle);
 676                         goto out;
 677                 }
 678
 679                 extend = true;
 680                 ext4_journal_stop(handle);
 681         }
 682
 683         ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
 684
 685         if (extend)
 686                 ret = ext4_handle_inode_extension(inode, offset, ret, count);
 687 out:
 688         inode_unlock(inode);
 689         if (ret > 0)
 690                 ret = generic_write_sync(iocb, ret);
 691         return ret;
 692 }
 693 #endif
 694
 695 static ssize_t
 696 ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
 697 {
 698         struct inode *inode = file_inode(iocb->ki_filp);
 699
 700         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 701                 return -EIO;
 702
 703 #ifdef CONFIG_FS_DAX
 704         if (IS_DAX(inode))
 705                 return ext4_dax_write_iter(iocb, from);
 706 #endif
 707         if (iocb->ki_flags & IOCB_DIRECT)
 708                 return ext4_dio_write_iter(iocb, from);
 709         else
 710                 return ext4_buffered_write_iter(iocb, from);
 711 }
 712
 713 #ifdef CONFIG_FS_DAX
 714 static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf,
 715                 enum page_entry_size pe_size)
 716 {
 717         int error = 0;
 718         vm_fault_t result;
 719         int retries = 0;
 720         handle_t *handle = NULL;
 721         struct inode *inode = file_inode(vmf->vma->vm_file);
 722         struct super_block *sb = inode->i_sb;
 723
 724         /*
 725          * We have to distinguish real writes from writes which will result in a
 726          * COW page; COW writes should *not* poke the journal (the file will not
 727          * be changed). Doing so would cause unintended failures when mounted
 728          * read-only.
 729          *
 730          * We check for VM_SHARED rather than vmf->cow_page since the latter is
 731          * unset for pe_size != PE_SIZE_PTE (i.e. only in do_cow_fault); for
 732          * other sizes, dax_iomap_fault will handle splitting / fallback so that
 733          * we eventually come back with a COW page.
 734          */
 735         bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
 736                 (vmf->vma->vm_flags & VM_SHARED);
 737         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
 738         pfn_t pfn;
 739
 740         if (write) {
 741                 sb_start_pagefault(sb);
 742                 file_update_time(vmf->vma->vm_file);
 743                 filemap_invalidate_lock_shared(mapping);
 744 retry:
 745                 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
 746                                                EXT4_DATA_TRANS_BLOCKS(sb));
 747                 if (IS_ERR(handle)) {
 748                         filemap_invalidate_unlock_shared(mapping);
 749                         sb_end_pagefault(sb);
 750                         return VM_FAULT_SIGBUS;
 751                 }
 752         } else {
 753                 filemap_invalidate_lock_shared(mapping);
 754         }
 755         result = dax_iomap_fault(vmf, pe_size, &pfn, &error, &ext4_iomap_ops);
 756         if (write) {
 757                 ext4_journal_stop(handle);
 758
 759                 if ((result & VM_FAULT_ERROR) && error == -ENOSPC &&
 760                     ext4_should_retry_alloc(sb, &retries))
 761                         goto retry;
 762                 /* Handling synchronous page fault? */
 763                 if (result & VM_FAULT_NEEDDSYNC)
 764                         result = dax_finish_sync_fault(vmf, pe_size, pfn);
 765                 filemap_invalidate_unlock_shared(mapping);
 766                 sb_end_pagefault(sb);
 767         } else {
 768                 filemap_invalidate_unlock_shared(mapping);
 769         }
 770
 771         return result;
 772 }
 773
 774 static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
 775 {
 776         return ext4_dax_huge_fault(vmf, PE_SIZE_PTE);
 777 }
 778
 779 static const struct vm_operations_struct ext4_dax_vm_ops = {
 780         .fault          = ext4_dax_fault,
 781         .huge_fault     = ext4_dax_huge_fault,
 782         .page_mkwrite   = ext4_dax_fault,
 783         .pfn_mkwrite    = ext4_dax_fault,
 784 };
 785 #else
 786 #define ext4_dax_vm_ops ext4_file_vm_ops
 787 #endif
 788
 789 static const struct vm_operations_struct ext4_file_vm_ops = {
 790         .fault          = filemap_fault,
 791         .map_pages      = filemap_map_pages,
 792         .page_mkwrite   = ext4_page_mkwrite,
 793 };
 794
 795 static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 796 {
 797         struct inode *inode = file->f_mapping->host;
 798         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 799         struct dax_device *dax_dev = sbi->s_daxdev;
 800
 801         if (unlikely(ext4_forced_shutdown(sbi)))
 802                 return -EIO;
 803
 804         /*
 805          * We don't support synchronous mappings for non-DAX files and
 806          * for DAX files if underneath dax_device is not synchronous.
 807          */
 808         if (!daxdev_mapping_supported(vma, dax_dev))
 809                 return -EOPNOTSUPP;
 810
 811         file_accessed(file);
 812         if (IS_DAX(file_inode(file))) {
 813                 vma->vm_ops = &ext4_dax_vm_ops;
 814                 vm_flags_set(vma, VM_HUGEPAGE);
 815         } else {
 816                 vma->vm_ops = &ext4_file_vm_ops;
 817         }
 818         return 0;
 819 }
 820
 821 static int ext4_sample_last_mounted(struct super_block *sb,
 822                                     struct vfsmount *mnt)
 823 {
 824         struct ext4_sb_info *sbi = EXT4_SB(sb);
 825         struct path path;
 826         char buf[64], *cp;
 827         handle_t *handle;
 828         int err;
 829
 830         if (likely(ext4_test_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED)))
 831                 return 0;
 832
 833         if (sb_rdonly(sb) || !sb_start_intwrite_trylock(sb))
 834                 return 0;
 835
 836         ext4_set_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED);
 837         /*
 838          * Sample where the filesystem has been mounted and
 839          * store it in the superblock for sysadmin convenience
 840          * when trying to sort through large numbers of block
 841          * devices or filesystem images.
 842          */
 843         memset(buf, 0, sizeof(buf));
 844         path.mnt = mnt;
 845         path.dentry = mnt->mnt_root;
 846         cp = d_path(&path, buf, sizeof(buf));
 847         err = 0;
 848         if (IS_ERR(cp))
 849                 goto out;
 850
 851         handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
 852         err = PTR_ERR(handle);
 853         if (IS_ERR(handle))
 854                 goto out;
 855         BUFFER_TRACE(sbi->s_sbh, "get_write_access");
 856         err = ext4_journal_get_write_access(handle, sb, sbi->s_sbh,
 857                                             EXT4_JTR_NONE);
 858         if (err)
 859                 goto out_journal;
 860         lock_buffer(sbi->s_sbh);
 861         strncpy(sbi->s_es->s_last_mounted, cp,
 862                 sizeof(sbi->s_es->s_last_mounted));
 863         ext4_superblock_csum_set(sb);
 864         unlock_buffer(sbi->s_sbh);
 865         ext4_handle_dirty_metadata(handle, NULL, sbi->s_sbh);
 866 out_journal:
 867         ext4_journal_stop(handle);
 868 out:
 869         sb_end_intwrite(sb);
 870         return err;
 871 }
 872
 873 static int ext4_file_open(struct inode *inode, struct file *filp)
 874 {
 875         int ret;
 876
 877         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 878                 return -EIO;
 879
 880         ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
 881         if (ret)
 882                 return ret;
 883
 884         ret = fscrypt_file_open(inode, filp);
 885         if (ret)
 886                 return ret;
 887
 888         ret = fsverity_file_open(inode, filp);
 889         if (ret)
 890                 return ret;
 891
 892         /*
 893          * Set up the jbd2_inode if we are opening the inode for
 894          * writing and the journal is present
 895          */
 896         if (filp->f_mode & FMODE_WRITE) {
 897                 ret = ext4_inode_attach_jinode(inode);
 898                 if (ret < 0)
 899                         return ret;
 900         }
 901
 902         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC |
 903                         FMODE_DIO_PARALLEL_WRITE;
 904         return dquot_file_open(inode, filp);
 905 }
 906
 907 /*
 908  * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
 909  * by calling generic_file_llseek_size() with the appropriate maxbytes
 910  * value for each.
 911  */
 912 loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
 913 {
 914         struct inode *inode = file->f_mapping->host;
 915         loff_t maxbytes;
 916
 917         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 918                 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
 919         else
 920                 maxbytes = inode->i_sb->s_maxbytes;
 921
 922         switch (whence) {
 923         default:
 924                 return generic_file_llseek_size(file, offset, whence,
 925                                                 maxbytes, i_size_read(inode));
 926         case SEEK_HOLE:
 927                 inode_lock_shared(inode);
 928                 offset = iomap_seek_hole(inode, offset,
 929                                          &ext4_iomap_report_ops);
 930                 inode_unlock_shared(inode);
 931                 break;
 932         case SEEK_DATA:
 933                 inode_lock_shared(inode);
 934                 offset = iomap_seek_data(inode, offset,
 935                                          &ext4_iomap_report_ops);
 936                 inode_unlock_shared(inode);
 937                 break;
 938         }
 939
 940         if (offset < 0)
 941                 return offset;
 942         return vfs_setpos(file, offset, maxbytes);
 943 }
 944
 945 const struct file_operations ext4_file_operations = {
 946         .llseek         = ext4_llseek,
 947         .read_iter      = ext4_file_read_iter,
 948         .write_iter     = ext4_file_write_iter,
 949         .iopoll         = iocb_bio_iopoll,
 950         .unlocked_ioctl = ext4_ioctl,
 951 #ifdef CONFIG_COMPAT
 952         .compat_ioctl   = ext4_compat_ioctl,
 953 #endif
 954         .mmap           = ext4_file_mmap,
 955         .mmap_supported_flags = MAP_SYNC,
 956         .open           = ext4_file_open,
 957         .release        = ext4_release_file,
 958         .fsync          = ext4_sync_file,
 959         .get_unmapped_area = thp_get_unmapped_area,
 960         .splice_read    = generic_file_splice_read,
 961         .splice_write   = iter_file_splice_write,
 962         .fallocate      = ext4_fallocate,
 963 };
 964
 965 const struct inode_operations ext4_file_inode_operations = {
 966         .setattr        = ext4_setattr,
 967         .getattr        = ext4_file_getattr,
 968         .listxattr      = ext4_listxattr,
 969         .get_inode_acl  = ext4_get_acl,
 970         .set_acl        = ext4_set_acl,
 971         .fiemap         = ext4_fiemap,
 972         .fileattr_get   = ext4_fileattr_get,
 973         .fileattr_set   = ext4_fileattr_set,
 974 };
 975