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 (card@masi.ibp.fr)
   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  *      (jj@sunsite.ms.mff.cuni.cz)
  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 static bool ext4_dio_supported(struct inode *inode)
  40 {
  41         if (IS_ENABLED(CONFIG_FS_ENCRYPTION) && IS_ENCRYPTED(inode))
  42                 return false;
  43         if (fsverity_active(inode))
  44                 return false;
  45         if (ext4_should_journal_data(inode))
  46                 return false;
  47         if (ext4_has_inline_data(inode))
  48                 return false;
  49         return true;
  50 }
  51
  52 static ssize_t ext4_dio_read_iter(struct kiocb *iocb, struct iov_iter *to)
  53 {
  54         ssize_t ret;
  55         struct inode *inode = file_inode(iocb->ki_filp);
  56
  57         if (iocb->ki_flags & IOCB_NOWAIT) {
  58                 if (!inode_trylock_shared(inode))
  59                         return -EAGAIN;
  60         } else {
  61                 inode_lock_shared(inode);
  62         }
  63
  64         if (!ext4_dio_supported(inode)) {
  65                 inode_unlock_shared(inode);
  66                 /*
  67                  * Fallback to buffered I/O if the operation being performed on
  68                  * the inode is not supported by direct I/O. The IOCB_DIRECT
  69                  * flag needs to be cleared here in order to ensure that the
  70                  * direct I/O path within generic_file_read_iter() is not
  71                  * taken.
  72                  */
  73                 iocb->ki_flags &= ~IOCB_DIRECT;
  74                 return generic_file_read_iter(iocb, to);
  75         }
  76
  77         ret = iomap_dio_rw(iocb, to, &ext4_iomap_ops, NULL,
  78                            is_sync_kiocb(iocb));
  79         inode_unlock_shared(inode);
  80
  81         file_accessed(iocb->ki_filp);
  82         return ret;
  83 }
  84
  85 #ifdef CONFIG_FS_DAX
  86 static ssize_t ext4_dax_read_iter(struct kiocb *iocb, struct iov_iter *to)
  87 {
  88         struct inode *inode = file_inode(iocb->ki_filp);
  89         ssize_t ret;
  90
  91         if (iocb->ki_flags & IOCB_NOWAIT) {
  92                 if (!inode_trylock_shared(inode))
  93                         return -EAGAIN;
  94         } else {
  95                 inode_lock_shared(inode);
  96         }
  97         /*
  98          * Recheck under inode lock - at this point we are sure it cannot
  99          * change anymore
 100          */
 101         if (!IS_DAX(inode)) {
 102                 inode_unlock_shared(inode);
 103                 /* Fallback to buffered IO in case we cannot support DAX */
 104                 return generic_file_read_iter(iocb, to);
 105         }
 106         ret = dax_iomap_rw(iocb, to, &ext4_iomap_ops);
 107         inode_unlock_shared(inode);
 108
 109         file_accessed(iocb->ki_filp);
 110         return ret;
 111 }
 112 #endif
 113
 114 static ssize_t ext4_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 115 {
 116         struct inode *inode = file_inode(iocb->ki_filp);
 117
 118         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 119                 return -EIO;
 120
 121         if (!iov_iter_count(to))
 122                 return 0; /* skip atime */
 123
 124 #ifdef CONFIG_FS_DAX
 125         if (IS_DAX(inode))
 126                 return ext4_dax_read_iter(iocb, to);
 127 #endif
 128         if (iocb->ki_flags & IOCB_DIRECT)
 129                 return ext4_dio_read_iter(iocb, to);
 130
 131         return generic_file_read_iter(iocb, to);
 132 }
 133
 134 /*
 135  * Called when an inode is released. Note that this is different
 136  * from ext4_file_open: open gets called at every open, but release
 137  * gets called only when /all/ the files are closed.
 138  */
 139 static int ext4_release_file(struct inode *inode, struct file *filp)
 140 {
 141         if (ext4_test_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE)) {
 142                 ext4_alloc_da_blocks(inode);
 143                 ext4_clear_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
 144         }
 145         /* if we are the last writer on the inode, drop the block reservation */
 146         if ((filp->f_mode & FMODE_WRITE) &&
 147                         (atomic_read(&inode->i_writecount) == 1) &&
 148                         !EXT4_I(inode)->i_reserved_data_blocks) {
 149                 down_write(&EXT4_I(inode)->i_data_sem);
 150                 ext4_discard_preallocations(inode, 0);
 151                 up_write(&EXT4_I(inode)->i_data_sem);
 152         }
 153         if (is_dx(inode) && filp->private_data)
 154                 ext4_htree_free_dir_info(filp->private_data);
 155
 156         return 0;
 157 }
 158
 159 /*
 160  * This tests whether the IO in question is block-aligned or not.
 161  * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
 162  * are converted to written only after the IO is complete.  Until they are
 163  * mapped, these blocks appear as holes, so dio_zero_block() will assume that
 164  * it needs to zero out portions of the start and/or end block.  If 2 AIO
 165  * threads are at work on the same unwritten block, they must be synchronized
 166  * or one thread will zero the other's data, causing corruption.
 167  */
 168 static bool
 169 ext4_unaligned_io(struct inode *inode, struct iov_iter *from, loff_t pos)
 170 {
 171         struct super_block *sb = inode->i_sb;
 172         unsigned long blockmask = sb->s_blocksize - 1;
 173
 174         if ((pos | iov_iter_alignment(from)) & blockmask)
 175                 return true;
 176
 177         return false;
 178 }
 179
 180 static bool
 181 ext4_extending_io(struct inode *inode, loff_t offset, size_t len)
 182 {
 183         if (offset + len > i_size_read(inode) ||
 184             offset + len > EXT4_I(inode)->i_disksize)
 185                 return true;
 186         return false;
 187 }
 188
 189 /* Is IO overwriting allocated and initialized blocks? */
 190 static bool ext4_overwrite_io(struct inode *inode, loff_t pos, loff_t len)
 191 {
 192         struct ext4_map_blocks map;
 193         unsigned int blkbits = inode->i_blkbits;
 194         int err, blklen;
 195
 196         if (pos + len > i_size_read(inode))
 197                 return false;
 198
 199         map.m_lblk = pos >> blkbits;
 200         map.m_len = EXT4_MAX_BLOCKS(len, pos, blkbits);
 201         blklen = map.m_len;
 202
 203         err = ext4_map_blocks(NULL, inode, &map, 0);
 204         /*
 205          * 'err==len' means that all of the blocks have been preallocated,
 206          * regardless of whether they have been initialized or not. To exclude
 207          * unwritten extents, we need to check m_flags.
 208          */
 209         return err == blklen && (map.m_flags & EXT4_MAP_MAPPED);
 210 }
 211
 212 static ssize_t ext4_generic_write_checks(struct kiocb *iocb,
 213                                          struct iov_iter *from)
 214 {
 215         struct inode *inode = file_inode(iocb->ki_filp);
 216         ssize_t ret;
 217
 218         if (unlikely(IS_IMMUTABLE(inode)))
 219                 return -EPERM;
 220
 221         ret = generic_write_checks(iocb, from);
 222         if (ret <= 0)
 223                 return ret;
 224
 225         /*
 226          * If we have encountered a bitmap-format file, the size limit
 227          * is smaller than s_maxbytes, which is for extent-mapped files.
 228          */
 229         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
 230                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 231
 232                 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes)
 233                         return -EFBIG;
 234                 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
 235         }
 236
 237         return iov_iter_count(from);
 238 }
 239
 240 static ssize_t ext4_write_checks(struct kiocb *iocb, struct iov_iter *from)
 241 {
 242         ssize_t ret, count;
 243
 244         count = ext4_generic_write_checks(iocb, from);
 245         if (count <= 0)
 246                 return count;
 247
 248         ret = file_modified(iocb->ki_filp);
 249         if (ret)
 250                 return ret;
 251         return count;
 252 }
 253
 254 static ssize_t ext4_buffered_write_iter(struct kiocb *iocb,
 255                                         struct iov_iter *from)
 256 {
 257         ssize_t ret;
 258         struct inode *inode = file_inode(iocb->ki_filp);
 259
 260         if (iocb->ki_flags & IOCB_NOWAIT)
 261                 return -EOPNOTSUPP;
 262
 263         inode_lock(inode);
 264         ret = ext4_write_checks(iocb, from);
 265         if (ret <= 0)
 266                 goto out;
 267
 268         current->backing_dev_info = inode_to_bdi(inode);
 269         ret = generic_perform_write(iocb->ki_filp, from, iocb->ki_pos);
 270         current->backing_dev_info = NULL;
 271
 272 out:
 273         inode_unlock(inode);
 274         if (likely(ret > 0)) {
 275                 iocb->ki_pos += ret;
 276                 ret = generic_write_sync(iocb, ret);
 277         }
 278
 279         return ret;
 280 }
 281
 282 static ssize_t ext4_handle_inode_extension(struct inode *inode, loff_t offset,
 283                                            ssize_t written, size_t count)
 284 {
 285         handle_t *handle;
 286         bool truncate = false;
 287         u8 blkbits = inode->i_blkbits;
 288         ext4_lblk_t written_blk, end_blk;
 289         int ret;
 290
 291         /*
 292          * Note that EXT4_I(inode)->i_disksize can get extended up to
 293          * inode->i_size while the I/O was running due to writeback of delalloc
 294          * blocks. But, the code in ext4_iomap_alloc() is careful to use
 295          * zeroed/unwritten extents if this is possible; thus we won't leave
 296          * uninitialized blocks in a file even if we didn't succeed in writing
 297          * as much as we intended.
 298          */
 299         WARN_ON_ONCE(i_size_read(inode) < EXT4_I(inode)->i_disksize);
 300         if (offset + count <= EXT4_I(inode)->i_disksize) {
 301                 /*
 302                  * We need to ensure that the inode is removed from the orphan
 303                  * list if it has been added prematurely, due to writeback of
 304                  * delalloc blocks.
 305                  */
 306                 if (!list_empty(&EXT4_I(inode)->i_orphan) && inode->i_nlink) {
 307                         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 308
 309                         if (IS_ERR(handle)) {
 310                                 ext4_orphan_del(NULL, inode);
 311                                 return PTR_ERR(handle);
 312                         }
 313
 314                         ext4_orphan_del(handle, inode);
 315                         ext4_journal_stop(handle);
 316                 }
 317
 318                 return written;
 319         }
 320
 321         if (written < 0)
 322                 goto truncate;
 323
 324         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 325         if (IS_ERR(handle)) {
 326                 written = PTR_ERR(handle);
 327                 goto truncate;
 328         }
 329
 330         if (ext4_update_inode_size(inode, offset + written)) {
 331                 ret = ext4_mark_inode_dirty(handle, inode);
 332                 if (unlikely(ret)) {
 333                         written = ret;
 334                         ext4_journal_stop(handle);
 335                         goto truncate;
 336                 }
 337         }
 338
 339         /*
 340          * We may need to truncate allocated but not written blocks beyond EOF.
 341          */
 342         written_blk = ALIGN(offset + written, 1 << blkbits);
 343         end_blk = ALIGN(offset + count, 1 << blkbits);
 344         if (written_blk < end_blk && ext4_can_truncate(inode))
 345                 truncate = true;
 346
 347         /*
 348          * Remove the inode from the orphan list if it has been extended and
 349          * everything went OK.
 350          */
 351         if (!truncate && inode->i_nlink)
 352                 ext4_orphan_del(handle, inode);
 353         ext4_journal_stop(handle);
 354
 355         if (truncate) {
 356 truncate:
 357                 ext4_truncate_failed_write(inode);
 358                 /*
 359                  * If the truncate operation failed early, then the inode may
 360                  * still be on the orphan list. In that case, we need to try
 361                  * remove the inode from the in-memory linked list.
 362                  */
 363                 if (inode->i_nlink)
 364                         ext4_orphan_del(NULL, inode);
 365         }
 366
 367         return written;
 368 }
 369
 370 static int ext4_dio_write_end_io(struct kiocb *iocb, ssize_t size,
 371                                  int error, unsigned int flags)
 372 {
 373         loff_t pos = iocb->ki_pos;
 374         struct inode *inode = file_inode(iocb->ki_filp);
 375
 376         if (error)
 377                 return error;
 378
 379         if (size && flags & IOMAP_DIO_UNWRITTEN) {
 380                 error = ext4_convert_unwritten_extents(NULL, inode, pos, size);
 381                 if (error < 0)
 382                         return error;
 383         }
 384         /*
 385          * If we are extending the file, we have to update i_size here before
 386          * page cache gets invalidated in iomap_dio_rw(). Otherwise racing
 387          * buffered reads could zero out too much from page cache pages. Update
 388          * of on-disk size will happen later in ext4_dio_write_iter() where
 389          * we have enough information to also perform orphan list handling etc.
 390          * Note that we perform all extending writes synchronously under
 391          * i_rwsem held exclusively so i_size update is safe here in that case.
 392          * If the write was not extending, we cannot see pos > i_size here
 393          * because operations reducing i_size like truncate wait for all
 394          * outstanding DIO before updating i_size.
 395          */
 396         pos += size;
 397         if (pos > i_size_read(inode))
 398                 i_size_write(inode, pos);
 399
 400         return 0;
 401 }
 402
 403 static const struct iomap_dio_ops ext4_dio_write_ops = {
 404         .end_io = ext4_dio_write_end_io,
 405 };
 406
 407 /*
 408  * The intention here is to start with shared lock acquired then see if any
 409  * condition requires an exclusive inode lock. If yes, then we restart the
 410  * whole operation by releasing the shared lock and acquiring exclusive lock.
 411  *
 412  * - For unaligned_io we never take shared lock as it may cause data corruption
 413  *   when two unaligned IO tries to modify the same block e.g. while zeroing.
 414  *
 415  * - For extending writes case we don't take the shared lock, since it requires
 416  *   updating inode i_disksize and/or orphan handling with exclusive lock.
 417  *
 418  * - shared locking will only be true mostly with overwrites. Otherwise we will
 419  *   switch to exclusive i_rwsem lock.
 420  */
 421 static ssize_t ext4_dio_write_checks(struct kiocb *iocb, struct iov_iter *from,
 422                                      bool *ilock_shared, bool *extend)
 423 {
 424         struct file *file = iocb->ki_filp;
 425         struct inode *inode = file_inode(file);
 426         loff_t offset;
 427         size_t count;
 428         ssize_t ret;
 429
 430 restart:
 431         ret = ext4_generic_write_checks(iocb, from);
 432         if (ret <= 0)
 433                 goto out;
 434
 435         offset = iocb->ki_pos;
 436         count = ret;
 437         if (ext4_extending_io(inode, offset, count))
 438                 *extend = true;
 439         /*
 440          * Determine whether the IO operation will overwrite allocated
 441          * and initialized blocks.
 442          * We need exclusive i_rwsem for changing security info
 443          * in file_modified().
 444          */
 445         if (*ilock_shared && (!IS_NOSEC(inode) || *extend ||
 446              !ext4_overwrite_io(inode, offset, count))) {
 447                 if (iocb->ki_flags & IOCB_NOWAIT) {
 448                         ret = -EAGAIN;
 449                         goto out;
 450                 }
 451                 inode_unlock_shared(inode);
 452                 *ilock_shared = false;
 453                 inode_lock(inode);
 454                 goto restart;
 455         }
 456
 457         ret = file_modified(file);
 458         if (ret < 0)
 459                 goto out;
 460
 461         return count;
 462 out:
 463         if (*ilock_shared)
 464                 inode_unlock_shared(inode);
 465         else
 466                 inode_unlock(inode);
 467         return ret;
 468 }
 469
 470 static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 471 {
 472         ssize_t ret;
 473         handle_t *handle;
 474         struct inode *inode = file_inode(iocb->ki_filp);
 475         loff_t offset = iocb->ki_pos;
 476         size_t count = iov_iter_count(from);
 477         const struct iomap_ops *iomap_ops = &ext4_iomap_ops;
 478         bool extend = false, unaligned_io = false;
 479         bool ilock_shared = true;
 480
 481         /*
 482          * We initially start with shared inode lock unless it is
 483          * unaligned IO which needs exclusive lock anyways.
 484          */
 485         if (ext4_unaligned_io(inode, from, offset)) {
 486                 unaligned_io = true;
 487                 ilock_shared = false;
 488         }
 489         /*
 490          * Quick check here without any i_rwsem lock to see if it is extending
 491          * IO. A more reliable check is done in ext4_dio_write_checks() with
 492          * proper locking in place.
 493          */
 494         if (offset + count > i_size_read(inode))
 495                 ilock_shared = false;
 496
 497         if (iocb->ki_flags & IOCB_NOWAIT) {
 498                 if (ilock_shared) {
 499                         if (!inode_trylock_shared(inode))
 500                                 return -EAGAIN;
 501                 } else {
 502                         if (!inode_trylock(inode))
 503                                 return -EAGAIN;
 504                 }
 505         } else {
 506                 if (ilock_shared)
 507                         inode_lock_shared(inode);
 508                 else
 509                         inode_lock(inode);
 510         }
 511
 512         /* Fallback to buffered I/O if the inode does not support direct I/O. */
 513         if (!ext4_dio_supported(inode)) {
 514                 if (ilock_shared)
 515                         inode_unlock_shared(inode);
 516                 else
 517                         inode_unlock(inode);
 518                 return ext4_buffered_write_iter(iocb, from);
 519         }
 520
 521         ret = ext4_dio_write_checks(iocb, from, &ilock_shared, &extend);
 522         if (ret <= 0)
 523                 return ret;
 524
 525         /* if we're going to block and IOCB_NOWAIT is set, return -EAGAIN */
 526         if ((iocb->ki_flags & IOCB_NOWAIT) && (unaligned_io || extend)) {
 527                 ret = -EAGAIN;
 528                 goto out;
 529         }
 530         /*
 531          * Make sure inline data cannot be created anymore since we are going
 532          * to allocate blocks for DIO. We know the inode does not have any
 533          * inline data now because ext4_dio_supported() checked for that.
 534          */
 535         ext4_clear_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
 536
 537         offset = iocb->ki_pos;
 538         count = ret;
 539
 540         /*
 541          * Unaligned direct IO must be serialized among each other as zeroing
 542          * of partial blocks of two competing unaligned IOs can result in data
 543          * corruption.
 544          *
 545          * So we make sure we don't allow any unaligned IO in flight.
 546          * For IOs where we need not wait (like unaligned non-AIO DIO),
 547          * below inode_dio_wait() may anyway become a no-op, since we start
 548          * with exclusive lock.
 549          */
 550         if (unaligned_io)
 551                 inode_dio_wait(inode);
 552
 553         if (extend) {
 554                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 555                 if (IS_ERR(handle)) {
 556                         ret = PTR_ERR(handle);
 557                         goto out;
 558                 }
 559
 560                 ret = ext4_orphan_add(handle, inode);
 561                 if (ret) {
 562                         ext4_journal_stop(handle);
 563                         goto out;
 564                 }
 565
 566                 ext4_journal_stop(handle);
 567         }
 568
 569         if (ilock_shared)
 570                 iomap_ops = &ext4_iomap_overwrite_ops;
 571         ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 572                            is_sync_kiocb(iocb) || unaligned_io || extend);
 573         if (ret == -ENOTBLK)
 574                 ret = 0;
 575
 576         if (extend)
 577                 ret = ext4_handle_inode_extension(inode, offset, ret, count);
 578
 579 out:
 580         if (ilock_shared)
 581                 inode_unlock_shared(inode);
 582         else
 583                 inode_unlock(inode);
 584
 585         if (ret >= 0 && iov_iter_count(from)) {
 586                 ssize_t err;
 587                 loff_t endbyte;
 588
 589                 offset = iocb->ki_pos;
 590                 err = ext4_buffered_write_iter(iocb, from);
 591                 if (err < 0)
 592                         return err;
 593
 594                 /*
 595                  * We need to ensure that the pages within the page cache for
 596                  * the range covered by this I/O are written to disk and
 597                  * invalidated. This is in attempt to preserve the expected
 598                  * direct I/O semantics in the case we fallback to buffered I/O
 599                  * to complete off the I/O request.
 600                  */
 601                 ret += err;
 602                 endbyte = offset + err - 1;
 603                 err = filemap_write_and_wait_range(iocb->ki_filp->f_mapping,
 604                                                    offset, endbyte);
 605                 if (!err)
 606                         invalidate_mapping_pages(iocb->ki_filp->f_mapping,
 607                                                  offset >> PAGE_SHIFT,
 608                                                  endbyte >> PAGE_SHIFT);
 609         }
 610
 611         return ret;
 612 }
 613
 614 #ifdef CONFIG_FS_DAX
 615 static ssize_t
 616 ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
 617 {
 618         ssize_t ret;
 619         size_t count;
 620         loff_t offset;
 621         handle_t *handle;
 622         bool extend = false;
 623         struct inode *inode = file_inode(iocb->ki_filp);
 624
 625         if (iocb->ki_flags & IOCB_NOWAIT) {
 626                 if (!inode_trylock(inode))
 627                         return -EAGAIN;
 628         } else {
 629                 inode_lock(inode);
 630         }
 631
 632         ret = ext4_write_checks(iocb, from);
 633         if (ret <= 0)
 634                 goto out;
 635
 636         offset = iocb->ki_pos;
 637         count = iov_iter_count(from);
 638
 639         if (offset + count > EXT4_I(inode)->i_disksize) {
 640                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 641                 if (IS_ERR(handle)) {
 642                         ret = PTR_ERR(handle);
 643                         goto out;
 644                 }
 645
 646                 ret = ext4_orphan_add(handle, inode);
 647                 if (ret) {
 648                         ext4_journal_stop(handle);
 649                         goto out;
 650                 }
 651
 652                 extend = true;
 653                 ext4_journal_stop(handle);
 654         }
 655
 656         ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
 657
 658         if (extend)
 659                 ret = ext4_handle_inode_extension(inode, offset, ret, count);
 660 out:
 661         inode_unlock(inode);
 662         if (ret > 0)
 663                 ret = generic_write_sync(iocb, ret);
 664         return ret;
 665 }
 666 #endif
 667
 668 static ssize_t
 669 ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
 670 {
 671         struct inode *inode = file_inode(iocb->ki_filp);
 672
 673         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 674                 return -EIO;
 675
 676 #ifdef CONFIG_FS_DAX
 677         if (IS_DAX(inode))
 678                 return ext4_dax_write_iter(iocb, from);
 679 #endif
 680         if (iocb->ki_flags & IOCB_DIRECT)
 681                 return ext4_dio_write_iter(iocb, from);
 682         else
 683                 return ext4_buffered_write_iter(iocb, from);
 684 }
 685
 686 #ifdef CONFIG_FS_DAX
 687 static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf,
 688                 enum page_entry_size pe_size)
 689 {
 690         int error = 0;
 691         vm_fault_t result;
 692         int retries = 0;
 693         handle_t *handle = NULL;
 694         struct inode *inode = file_inode(vmf->vma->vm_file);
 695         struct super_block *sb = inode->i_sb;
 696
 697         /*
 698          * We have to distinguish real writes from writes which will result in a
 699          * COW page; COW writes should *not* poke the journal (the file will not
 700          * be changed). Doing so would cause unintended failures when mounted
 701          * read-only.
 702          *
 703          * We check for VM_SHARED rather than vmf->cow_page since the latter is
 704          * unset for pe_size != PE_SIZE_PTE (i.e. only in do_cow_fault); for
 705          * other sizes, dax_iomap_fault will handle splitting / fallback so that
 706          * we eventually come back with a COW page.
 707          */
 708         bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
 709                 (vmf->vma->vm_flags & VM_SHARED);
 710         pfn_t pfn;
 711
 712         if (write) {
 713                 sb_start_pagefault(sb);
 714                 file_update_time(vmf->vma->vm_file);
 715                 down_read(&EXT4_I(inode)->i_mmap_sem);
 716 retry:
 717                 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
 718                                                EXT4_DATA_TRANS_BLOCKS(sb));
 719                 if (IS_ERR(handle)) {
 720                         up_read(&EXT4_I(inode)->i_mmap_sem);
 721                         sb_end_pagefault(sb);
 722                         return VM_FAULT_SIGBUS;
 723                 }
 724         } else {
 725                 down_read(&EXT4_I(inode)->i_mmap_sem);
 726         }
 727         result = dax_iomap_fault(vmf, pe_size, &pfn, &error, &ext4_iomap_ops);
 728         if (write) {
 729                 ext4_journal_stop(handle);
 730
 731                 if ((result & VM_FAULT_ERROR) && error == -ENOSPC &&
 732                     ext4_should_retry_alloc(sb, &retries))
 733                         goto retry;
 734                 /* Handling synchronous page fault? */
 735                 if (result & VM_FAULT_NEEDDSYNC)
 736                         result = dax_finish_sync_fault(vmf, pe_size, pfn);
 737                 up_read(&EXT4_I(inode)->i_mmap_sem);
 738                 sb_end_pagefault(sb);
 739         } else {
 740                 up_read(&EXT4_I(inode)->i_mmap_sem);
 741         }
 742
 743         return result;
 744 }
 745
 746 static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
 747 {
 748         return ext4_dax_huge_fault(vmf, PE_SIZE_PTE);
 749 }
 750
 751 static const struct vm_operations_struct ext4_dax_vm_ops = {
 752         .fault          = ext4_dax_fault,
 753         .huge_fault     = ext4_dax_huge_fault,
 754         .page_mkwrite   = ext4_dax_fault,
 755         .pfn_mkwrite    = ext4_dax_fault,
 756 };
 757 #else
 758 #define ext4_dax_vm_ops ext4_file_vm_ops
 759 #endif
 760
 761 static const struct vm_operations_struct ext4_file_vm_ops = {
 762         .fault          = ext4_filemap_fault,
 763         .map_pages      = filemap_map_pages,
 764         .page_mkwrite   = ext4_page_mkwrite,
 765 };
 766
 767 static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 768 {
 769         struct inode *inode = file->f_mapping->host;
 770         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 771         struct dax_device *dax_dev = sbi->s_daxdev;
 772
 773         if (unlikely(ext4_forced_shutdown(sbi)))
 774                 return -EIO;
 775
 776         /*
 777          * We don't support synchronous mappings for non-DAX files and
 778          * for DAX files if underneath dax_device is not synchronous.
 779          */
 780         if (!daxdev_mapping_supported(vma, dax_dev))
 781                 return -EOPNOTSUPP;
 782
 783         file_accessed(file);
 784         if (IS_DAX(file_inode(file))) {
 785                 vma->vm_ops = &ext4_dax_vm_ops;
 786                 vma->vm_flags |= VM_HUGEPAGE;
 787         } else {
 788                 vma->vm_ops = &ext4_file_vm_ops;
 789         }
 790         return 0;
 791 }
 792
 793 static int ext4_sample_last_mounted(struct super_block *sb,
 794                                     struct vfsmount *mnt)
 795 {
 796         struct ext4_sb_info *sbi = EXT4_SB(sb);
 797         struct path path;
 798         char buf[64], *cp;
 799         handle_t *handle;
 800         int err;
 801
 802         if (likely(ext4_test_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED)))
 803                 return 0;
 804
 805         if (sb_rdonly(sb) || !sb_start_intwrite_trylock(sb))
 806                 return 0;
 807
 808         ext4_set_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED);
 809         /*
 810          * Sample where the filesystem has been mounted and
 811          * store it in the superblock for sysadmin convenience
 812          * when trying to sort through large numbers of block
 813          * devices or filesystem images.
 814          */
 815         memset(buf, 0, sizeof(buf));
 816         path.mnt = mnt;
 817         path.dentry = mnt->mnt_root;
 818         cp = d_path(&path, buf, sizeof(buf));
 819         err = 0;
 820         if (IS_ERR(cp))
 821                 goto out;
 822
 823         handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
 824         err = PTR_ERR(handle);
 825         if (IS_ERR(handle))
 826                 goto out;
 827         BUFFER_TRACE(sbi->s_sbh, "get_write_access");
 828         err = ext4_journal_get_write_access(handle, sbi->s_sbh);
 829         if (err)
 830                 goto out_journal;
 831         strncpy(sbi->s_es->s_last_mounted, cp,
 832                 sizeof(sbi->s_es->s_last_mounted));
 833         ext4_handle_dirty_super(handle, sb);
 834 out_journal:
 835         ext4_journal_stop(handle);
 836 out:
 837         sb_end_intwrite(sb);
 838         return err;
 839 }
 840
 841 static int ext4_file_open(struct inode *inode, struct file *filp)
 842 {
 843         int ret;
 844
 845         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 846                 return -EIO;
 847
 848         ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
 849         if (ret)
 850                 return ret;
 851
 852         ret = fscrypt_file_open(inode, filp);
 853         if (ret)
 854                 return ret;
 855
 856         ret = fsverity_file_open(inode, filp);
 857         if (ret)
 858                 return ret;
 859
 860         /*
 861          * Set up the jbd2_inode if we are opening the inode for
 862          * writing and the journal is present
 863          */
 864         if (filp->f_mode & FMODE_WRITE) {
 865                 ret = ext4_inode_attach_jinode(inode);
 866                 if (ret < 0)
 867                         return ret;
 868         }
 869
 870         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
 871         return dquot_file_open(inode, filp);
 872 }
 873
 874 /*
 875  * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
 876  * by calling generic_file_llseek_size() with the appropriate maxbytes
 877  * value for each.
 878  */
 879 loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
 880 {
 881         struct inode *inode = file->f_mapping->host;
 882         loff_t maxbytes;
 883
 884         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 885                 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
 886         else
 887                 maxbytes = inode->i_sb->s_maxbytes;
 888
 889         switch (whence) {
 890         default:
 891                 return generic_file_llseek_size(file, offset, whence,
 892                                                 maxbytes, i_size_read(inode));
 893         case SEEK_HOLE:
 894                 inode_lock_shared(inode);
 895                 offset = iomap_seek_hole(inode, offset,
 896                                          &ext4_iomap_report_ops);
 897                 inode_unlock_shared(inode);
 898                 break;
 899         case SEEK_DATA:
 900                 inode_lock_shared(inode);
 901                 offset = iomap_seek_data(inode, offset,
 902                                          &ext4_iomap_report_ops);
 903                 inode_unlock_shared(inode);
 904                 break;
 905         }
 906
 907         if (offset < 0)
 908                 return offset;
 909         return vfs_setpos(file, offset, maxbytes);
 910 }
 911
 912 const struct file_operations ext4_file_operations = {
 913         .llseek         = ext4_llseek,
 914         .read_iter      = ext4_file_read_iter,
 915         .write_iter     = ext4_file_write_iter,
 916         .iopoll         = iomap_dio_iopoll,
 917         .unlocked_ioctl = ext4_ioctl,
 918 #ifdef CONFIG_COMPAT
 919         .compat_ioctl   = ext4_compat_ioctl,
 920 #endif
 921         .mmap           = ext4_file_mmap,
 922         .mmap_supported_flags = MAP_SYNC,
 923         .open           = ext4_file_open,
 924         .release        = ext4_release_file,
 925         .fsync          = ext4_sync_file,
 926         .get_unmapped_area = thp_get_unmapped_area,
 927         .splice_read    = generic_file_splice_read,
 928         .splice_write   = iter_file_splice_write,
 929         .fallocate      = ext4_fallocate,
 930 };
 931
 932 const struct inode_operations ext4_file_inode_operations = {
 933         .setattr        = ext4_setattr,
 934         .getattr        = ext4_file_getattr,
 935         .listxattr      = ext4_listxattr,
 936         .get_acl        = ext4_get_acl,
 937         .set_acl        = ext4_set_acl,
 938         .fiemap         = ext4_fiemap,
 939 };
 940