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(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 static ssize_t ext4_file_splice_read(struct file *in, loff_t *ppos,
 151                                      struct pipe_inode_info *pipe,
 152                                      size_t len, unsigned int flags)
 153 {
 154         struct inode *inode = file_inode(in);
 155
 156         if (unlikely(ext4_forced_shutdown(inode->i_sb)))
 157                 return -EIO;
 158         return filemap_splice_read(in, ppos, pipe, len, flags);
 159 }
 160
 161 /*
 162  * Called when an inode is released. Note that this is different
 163  * from ext4_file_open: open gets called at every open, but release
 164  * gets called only when /all/ the files are closed.
 165  */
 166 static int ext4_release_file(struct inode *inode, struct file *filp)
 167 {
 168         if (ext4_test_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE)) {
 169                 ext4_alloc_da_blocks(inode);
 170                 ext4_clear_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
 171         }
 172         /* if we are the last writer on the inode, drop the block reservation */
 173         if ((filp->f_mode & FMODE_WRITE) &&
 174                         (atomic_read(&inode->i_writecount) == 1) &&
 175                         !EXT4_I(inode)->i_reserved_data_blocks) {
 176                 down_write(&EXT4_I(inode)->i_data_sem);
 177                 ext4_discard_preallocations(inode);
 178                 up_write(&EXT4_I(inode)->i_data_sem);
 179         }
 180         if (is_dx(inode) && filp->private_data)
 181                 ext4_htree_free_dir_info(filp->private_data);
 182
 183         return 0;
 184 }
 185
 186 /*
 187  * This tests whether the IO in question is block-aligned or not.
 188  * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
 189  * are converted to written only after the IO is complete.  Until they are
 190  * mapped, these blocks appear as holes, so dio_zero_block() will assume that
 191  * it needs to zero out portions of the start and/or end block.  If 2 AIO
 192  * threads are at work on the same unwritten block, they must be synchronized
 193  * or one thread will zero the other's data, causing corruption.
 194  */
 195 static bool
 196 ext4_unaligned_io(struct inode *inode, struct iov_iter *from, loff_t pos)
 197 {
 198         struct super_block *sb = inode->i_sb;
 199         unsigned long blockmask = sb->s_blocksize - 1;
 200
 201         if ((pos | iov_iter_alignment(from)) & blockmask)
 202                 return true;
 203
 204         return false;
 205 }
 206
 207 static bool
 208 ext4_extending_io(struct inode *inode, loff_t offset, size_t len)
 209 {
 210         if (offset + len > i_size_read(inode) ||
 211             offset + len > EXT4_I(inode)->i_disksize)
 212                 return true;
 213         return false;
 214 }
 215
 216 /* Is IO overwriting allocated or initialized blocks? */
 217 static bool ext4_overwrite_io(struct inode *inode,
 218                               loff_t pos, loff_t len, bool *unwritten)
 219 {
 220         struct ext4_map_blocks map;
 221         unsigned int blkbits = inode->i_blkbits;
 222         int err, blklen;
 223
 224         if (pos + len > i_size_read(inode))
 225                 return false;
 226
 227         map.m_lblk = pos >> blkbits;
 228         map.m_len = EXT4_MAX_BLOCKS(len, pos, blkbits);
 229         blklen = map.m_len;
 230
 231         err = ext4_map_blocks(NULL, inode, &map, 0);
 232         if (err != blklen)
 233                 return false;
 234         /*
 235          * 'err==len' means that all of the blocks have been preallocated,
 236          * regardless of whether they have been initialized or not. We need to
 237          * check m_flags to distinguish the unwritten extents.
 238          */
 239         *unwritten = !(map.m_flags & EXT4_MAP_MAPPED);
 240         return true;
 241 }
 242
 243 static ssize_t ext4_generic_write_checks(struct kiocb *iocb,
 244                                          struct iov_iter *from)
 245 {
 246         struct inode *inode = file_inode(iocb->ki_filp);
 247         ssize_t ret;
 248
 249         if (unlikely(IS_IMMUTABLE(inode)))
 250                 return -EPERM;
 251
 252         ret = generic_write_checks(iocb, from);
 253         if (ret <= 0)
 254                 return ret;
 255
 256         /*
 257          * If we have encountered a bitmap-format file, the size limit
 258          * is smaller than s_maxbytes, which is for extent-mapped files.
 259          */
 260         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
 261                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 262
 263                 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes)
 264                         return -EFBIG;
 265                 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
 266         }
 267
 268         return iov_iter_count(from);
 269 }
 270
 271 static ssize_t ext4_write_checks(struct kiocb *iocb, struct iov_iter *from)
 272 {
 273         ssize_t ret, count;
 274
 275         count = ext4_generic_write_checks(iocb, from);
 276         if (count <= 0)
 277                 return count;
 278
 279         ret = file_modified(iocb->ki_filp);
 280         if (ret)
 281                 return ret;
 282         return count;
 283 }
 284
 285 static ssize_t ext4_buffered_write_iter(struct kiocb *iocb,
 286                                         struct iov_iter *from)
 287 {
 288         ssize_t ret;
 289         struct inode *inode = file_inode(iocb->ki_filp);
 290
 291         if (iocb->ki_flags & IOCB_NOWAIT)
 292                 return -EOPNOTSUPP;
 293
 294         inode_lock(inode);
 295         ret = ext4_write_checks(iocb, from);
 296         if (ret <= 0)
 297                 goto out;
 298
 299         ret = generic_perform_write(iocb, from);
 300
 301 out:
 302         inode_unlock(inode);
 303         if (unlikely(ret <= 0))
 304                 return ret;
 305         return generic_write_sync(iocb, ret);
 306 }
 307
 308 static ssize_t ext4_handle_inode_extension(struct inode *inode, loff_t offset,
 309                                            ssize_t count)
 310 {
 311         handle_t *handle;
 312
 313         lockdep_assert_held_write(&inode->i_rwsem);
 314         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 315         if (IS_ERR(handle))
 316                 return PTR_ERR(handle);
 317
 318         if (ext4_update_inode_size(inode, offset + count)) {
 319                 int ret = ext4_mark_inode_dirty(handle, inode);
 320                 if (unlikely(ret)) {
 321                         ext4_journal_stop(handle);
 322                         return ret;
 323                 }
 324         }
 325
 326         if (inode->i_nlink)
 327                 ext4_orphan_del(handle, inode);
 328         ext4_journal_stop(handle);
 329
 330         return count;
 331 }
 332
 333 /*
 334  * Clean up the inode after DIO or DAX extending write has completed and the
 335  * inode size has been updated using ext4_handle_inode_extension().
 336  */
 337 static void ext4_inode_extension_cleanup(struct inode *inode, ssize_t count)
 338 {
 339         lockdep_assert_held_write(&inode->i_rwsem);
 340         if (count < 0) {
 341                 ext4_truncate_failed_write(inode);
 342                 /*
 343                  * If the truncate operation failed early, then the inode may
 344                  * still be on the orphan list. In that case, we need to try
 345                  * remove the inode from the in-memory linked list.
 346                  */
 347                 if (inode->i_nlink)
 348                         ext4_orphan_del(NULL, inode);
 349                 return;
 350         }
 351         /*
 352          * If i_disksize got extended either due to writeback of delalloc
 353          * blocks or extending truncate while the DIO was running we could fail
 354          * to cleanup the orphan list in ext4_handle_inode_extension(). Do it
 355          * now.
 356          */
 357         if (!list_empty(&EXT4_I(inode)->i_orphan) && inode->i_nlink) {
 358                 handle_t *handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 359
 360                 if (IS_ERR(handle)) {
 361                         /*
 362                          * The write has successfully completed. Not much to
 363                          * do with the error here so just cleanup the orphan
 364                          * list and hope for the best.
 365                          */
 366                         ext4_orphan_del(NULL, inode);
 367                         return;
 368                 }
 369                 ext4_orphan_del(handle, inode);
 370                 ext4_journal_stop(handle);
 371         }
 372 }
 373
 374 static int ext4_dio_write_end_io(struct kiocb *iocb, ssize_t size,
 375                                  int error, unsigned int flags)
 376 {
 377         loff_t pos = iocb->ki_pos;
 378         struct inode *inode = file_inode(iocb->ki_filp);
 379
 380         if (!error && size && flags & IOMAP_DIO_UNWRITTEN)
 381                 error = ext4_convert_unwritten_extents(NULL, inode, pos, size);
 382         if (error)
 383                 return error;
 384         /*
 385          * Note that EXT4_I(inode)->i_disksize can get extended up to
 386          * inode->i_size while the I/O was running due to writeback of delalloc
 387          * blocks. But the code in ext4_iomap_alloc() is careful to use
 388          * zeroed/unwritten extents if this is possible; thus we won't leave
 389          * uninitialized blocks in a file even if we didn't succeed in writing
 390          * as much as we intended. Also we can race with truncate or write
 391          * expanding the file so we have to be a bit careful here.
 392          */
 393         if (pos + size <= READ_ONCE(EXT4_I(inode)->i_disksize) &&
 394             pos + size <= i_size_read(inode))
 395                 return size;
 396         return ext4_handle_inode_extension(inode, pos, size);
 397 }
 398
 399 static const struct iomap_dio_ops ext4_dio_write_ops = {
 400         .end_io = ext4_dio_write_end_io,
 401 };
 402
 403 /*
 404  * The intention here is to start with shared lock acquired then see if any
 405  * condition requires an exclusive inode lock. If yes, then we restart the
 406  * whole operation by releasing the shared lock and acquiring exclusive lock.
 407  *
 408  * - For unaligned_io we never take shared lock as it may cause data corruption
 409  *   when two unaligned IO tries to modify the same block e.g. while zeroing.
 410  *
 411  * - For extending writes case we don't take the shared lock, since it requires
 412  *   updating inode i_disksize and/or orphan handling with exclusive lock.
 413  *
 414  * - shared locking will only be true mostly with overwrites, including
 415  *   initialized blocks and unwritten blocks. For overwrite unwritten blocks
 416  *   we protect splitting extents by i_data_sem in ext4_inode_info, so we can
 417  *   also release exclusive i_rwsem lock.
 418  *
 419  * - Otherwise we will 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                                      bool *unwritten, int *dio_flags)
 424 {
 425         struct file *file = iocb->ki_filp;
 426         struct inode *inode = file_inode(file);
 427         loff_t offset;
 428         size_t count;
 429         ssize_t ret;
 430         bool overwrite, unaligned_io;
 431
 432 restart:
 433         ret = ext4_generic_write_checks(iocb, from);
 434         if (ret <= 0)
 435                 goto out;
 436
 437         offset = iocb->ki_pos;
 438         count = ret;
 439
 440         unaligned_io = ext4_unaligned_io(inode, from, offset);
 441         *extend = ext4_extending_io(inode, offset, count);
 442         overwrite = ext4_overwrite_io(inode, offset, count, unwritten);
 443
 444         /*
 445          * Determine whether we need to upgrade to an exclusive lock. This is
 446          * required to change security info in file_modified(), for extending
 447          * I/O, any form of non-overwrite I/O, and unaligned I/O to unwritten
 448          * extents (as partial block zeroing may be required).
 449          *
 450          * Note that unaligned writes are allowed under shared lock so long as
 451          * they are pure overwrites. Otherwise, concurrent unaligned writes risk
 452          * data corruption due to partial block zeroing in the dio layer, and so
 453          * the I/O must occur exclusively.
 454          */
 455         if (*ilock_shared &&
 456             ((!IS_NOSEC(inode) || *extend || !overwrite ||
 457              (unaligned_io && *unwritten)))) {
 458                 if (iocb->ki_flags & IOCB_NOWAIT) {
 459                         ret = -EAGAIN;
 460                         goto out;
 461                 }
 462                 inode_unlock_shared(inode);
 463                 *ilock_shared = false;
 464                 inode_lock(inode);
 465                 goto restart;
 466         }
 467
 468         /*
 469          * Now that locking is settled, determine dio flags and exclusivity
 470          * requirements. We don't use DIO_OVERWRITE_ONLY because we enforce
 471          * behavior already. The inode lock is already held exclusive if the
 472          * write is non-overwrite or extending, so drain all outstanding dio and
 473          * set the force wait dio flag.
 474          */
 475         if (!*ilock_shared && (unaligned_io || *extend)) {
 476                 if (iocb->ki_flags & IOCB_NOWAIT) {
 477                         ret = -EAGAIN;
 478                         goto out;
 479                 }
 480                 if (unaligned_io && (!overwrite || *unwritten))
 481                         inode_dio_wait(inode);
 482                 *dio_flags = IOMAP_DIO_FORCE_WAIT;
 483         }
 484
 485         ret = file_modified(file);
 486         if (ret < 0)
 487                 goto out;
 488
 489         return count;
 490 out:
 491         if (*ilock_shared)
 492                 inode_unlock_shared(inode);
 493         else
 494                 inode_unlock(inode);
 495         return ret;
 496 }
 497
 498 static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 499 {
 500         ssize_t ret;
 501         handle_t *handle;
 502         struct inode *inode = file_inode(iocb->ki_filp);
 503         loff_t offset = iocb->ki_pos;
 504         size_t count = iov_iter_count(from);
 505         const struct iomap_ops *iomap_ops = &ext4_iomap_ops;
 506         bool extend = false, unwritten = false;
 507         bool ilock_shared = true;
 508         int dio_flags = 0;
 509
 510         /*
 511          * Quick check here without any i_rwsem lock to see if it is extending
 512          * IO. A more reliable check is done in ext4_dio_write_checks() with
 513          * proper locking in place.
 514          */
 515         if (offset + count > i_size_read(inode))
 516                 ilock_shared = false;
 517
 518         if (iocb->ki_flags & IOCB_NOWAIT) {
 519                 if (ilock_shared) {
 520                         if (!inode_trylock_shared(inode))
 521                                 return -EAGAIN;
 522                 } else {
 523                         if (!inode_trylock(inode))
 524                                 return -EAGAIN;
 525                 }
 526         } else {
 527                 if (ilock_shared)
 528                         inode_lock_shared(inode);
 529                 else
 530                         inode_lock(inode);
 531         }
 532
 533         /* Fallback to buffered I/O if the inode does not support direct I/O. */
 534         if (!ext4_should_use_dio(iocb, from)) {
 535                 if (ilock_shared)
 536                         inode_unlock_shared(inode);
 537                 else
 538                         inode_unlock(inode);
 539                 return ext4_buffered_write_iter(iocb, from);
 540         }
 541
 542         /*
 543          * Prevent inline data from being created since we are going to allocate
 544          * blocks for DIO. We know the inode does not currently have inline data
 545          * because ext4_should_use_dio() checked for it, but we have to clear
 546          * the state flag before the write checks because a lock cycle could
 547          * introduce races with other writers.
 548          */
 549         ext4_clear_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
 550
 551         ret = ext4_dio_write_checks(iocb, from, &ilock_shared, &extend,
 552                                     &unwritten, &dio_flags);
 553         if (ret <= 0)
 554                 return ret;
 555
 556         offset = iocb->ki_pos;
 557         count = ret;
 558
 559         if (extend) {
 560                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 561                 if (IS_ERR(handle)) {
 562                         ret = PTR_ERR(handle);
 563                         goto out;
 564                 }
 565
 566                 ret = ext4_orphan_add(handle, inode);
 567                 if (ret) {
 568                         ext4_journal_stop(handle);
 569                         goto out;
 570                 }
 571
 572                 ext4_journal_stop(handle);
 573         }
 574
 575         if (ilock_shared && !unwritten)
 576                 iomap_ops = &ext4_iomap_overwrite_ops;
 577         ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 578                            dio_flags, NULL, 0);
 579         if (ret == -ENOTBLK)
 580                 ret = 0;
 581         if (extend) {
 582                 /*
 583                  * We always perform extending DIO write synchronously so by
 584                  * now the IO is completed and ext4_handle_inode_extension()
 585                  * was called. Cleanup the inode in case of error or race with
 586                  * writeback of delalloc blocks.
 587                  */
 588                 WARN_ON_ONCE(ret == -EIOCBQUEUED);
 589                 ext4_inode_extension_cleanup(inode, ret);
 590         }
 591
 592 out:
 593         if (ilock_shared)
 594                 inode_unlock_shared(inode);
 595         else
 596                 inode_unlock(inode);
 597
 598         if (ret >= 0 && iov_iter_count(from)) {
 599                 ssize_t err;
 600                 loff_t endbyte;
 601
 602                 offset = iocb->ki_pos;
 603                 err = ext4_buffered_write_iter(iocb, from);
 604                 if (err < 0)
 605                         return err;
 606
 607                 /*
 608                  * We need to ensure that the pages within the page cache for
 609                  * the range covered by this I/O are written to disk and
 610                  * invalidated. This is in attempt to preserve the expected
 611                  * direct I/O semantics in the case we fallback to buffered I/O
 612                  * to complete off the I/O request.
 613                  */
 614                 ret += err;
 615                 endbyte = offset + err - 1;
 616                 err = filemap_write_and_wait_range(iocb->ki_filp->f_mapping,
 617                                                    offset, endbyte);
 618                 if (!err)
 619                         invalidate_mapping_pages(iocb->ki_filp->f_mapping,
 620                                                  offset >> PAGE_SHIFT,
 621                                                  endbyte >> PAGE_SHIFT);
 622         }
 623
 624         return ret;
 625 }
 626
 627 #ifdef CONFIG_FS_DAX
 628 static ssize_t
 629 ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
 630 {
 631         ssize_t ret;
 632         size_t count;
 633         loff_t offset;
 634         handle_t *handle;
 635         bool extend = false;
 636         struct inode *inode = file_inode(iocb->ki_filp);
 637
 638         if (iocb->ki_flags & IOCB_NOWAIT) {
 639                 if (!inode_trylock(inode))
 640                         return -EAGAIN;
 641         } else {
 642                 inode_lock(inode);
 643         }
 644
 645         ret = ext4_write_checks(iocb, from);
 646         if (ret <= 0)
 647                 goto out;
 648
 649         offset = iocb->ki_pos;
 650         count = iov_iter_count(from);
 651
 652         if (offset + count > EXT4_I(inode)->i_disksize) {
 653                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 654                 if (IS_ERR(handle)) {
 655                         ret = PTR_ERR(handle);
 656                         goto out;
 657                 }
 658
 659                 ret = ext4_orphan_add(handle, inode);
 660                 if (ret) {
 661                         ext4_journal_stop(handle);
 662                         goto out;
 663                 }
 664
 665                 extend = true;
 666                 ext4_journal_stop(handle);
 667         }
 668
 669         ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
 670
 671         if (extend) {
 672                 ret = ext4_handle_inode_extension(inode, offset, ret);
 673                 ext4_inode_extension_cleanup(inode, ret);
 674         }
 675 out:
 676         inode_unlock(inode);
 677         if (ret > 0)
 678                 ret = generic_write_sync(iocb, ret);
 679         return ret;
 680 }
 681 #endif
 682
 683 static ssize_t
 684 ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
 685 {
 686         struct inode *inode = file_inode(iocb->ki_filp);
 687
 688         if (unlikely(ext4_forced_shutdown(inode->i_sb)))
 689                 return -EIO;
 690
 691 #ifdef CONFIG_FS_DAX
 692         if (IS_DAX(inode))
 693                 return ext4_dax_write_iter(iocb, from);
 694 #endif
 695         if (iocb->ki_flags & IOCB_DIRECT)
 696                 return ext4_dio_write_iter(iocb, from);
 697         else
 698                 return ext4_buffered_write_iter(iocb, from);
 699 }
 700
 701 #ifdef CONFIG_FS_DAX
 702 static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf, unsigned int order)
 703 {
 704         int error = 0;
 705         vm_fault_t result;
 706         int retries = 0;
 707         handle_t *handle = NULL;
 708         struct inode *inode = file_inode(vmf->vma->vm_file);
 709         struct super_block *sb = inode->i_sb;
 710
 711         /*
 712          * We have to distinguish real writes from writes which will result in a
 713          * COW page; COW writes should *not* poke the journal (the file will not
 714          * be changed). Doing so would cause unintended failures when mounted
 715          * read-only.
 716          *
 717          * We check for VM_SHARED rather than vmf->cow_page since the latter is
 718          * unset for order != 0 (i.e. only in do_cow_fault); for
 719          * other sizes, dax_iomap_fault will handle splitting / fallback so that
 720          * we eventually come back with a COW page.
 721          */
 722         bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
 723                 (vmf->vma->vm_flags & VM_SHARED);
 724         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
 725         pfn_t pfn;
 726
 727         if (write) {
 728                 sb_start_pagefault(sb);
 729                 file_update_time(vmf->vma->vm_file);
 730                 filemap_invalidate_lock_shared(mapping);
 731 retry:
 732                 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
 733                                                EXT4_DATA_TRANS_BLOCKS(sb));
 734                 if (IS_ERR(handle)) {
 735                         filemap_invalidate_unlock_shared(mapping);
 736                         sb_end_pagefault(sb);
 737                         return VM_FAULT_SIGBUS;
 738                 }
 739         } else {
 740                 filemap_invalidate_lock_shared(mapping);
 741         }
 742         result = dax_iomap_fault(vmf, order, &pfn, &error, &ext4_iomap_ops);
 743         if (write) {
 744                 ext4_journal_stop(handle);
 745
 746                 if ((result & VM_FAULT_ERROR) && error == -ENOSPC &&
 747                     ext4_should_retry_alloc(sb, &retries))
 748                         goto retry;
 749                 /* Handling synchronous page fault? */
 750                 if (result & VM_FAULT_NEEDDSYNC)
 751                         result = dax_finish_sync_fault(vmf, order, pfn);
 752                 filemap_invalidate_unlock_shared(mapping);
 753                 sb_end_pagefault(sb);
 754         } else {
 755                 filemap_invalidate_unlock_shared(mapping);
 756         }
 757
 758         return result;
 759 }
 760
 761 static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
 762 {
 763         return ext4_dax_huge_fault(vmf, 0);
 764 }
 765
 766 static const struct vm_operations_struct ext4_dax_vm_ops = {
 767         .fault          = ext4_dax_fault,
 768         .huge_fault     = ext4_dax_huge_fault,
 769         .page_mkwrite   = ext4_dax_fault,
 770         .pfn_mkwrite    = ext4_dax_fault,
 771 };
 772 #else
 773 #define ext4_dax_vm_ops ext4_file_vm_ops
 774 #endif
 775
 776 static const struct vm_operations_struct ext4_file_vm_ops = {
 777         .fault          = filemap_fault,
 778         .map_pages      = filemap_map_pages,
 779         .page_mkwrite   = ext4_page_mkwrite,
 780 };
 781
 782 static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 783 {
 784         struct inode *inode = file->f_mapping->host;
 785         struct dax_device *dax_dev = EXT4_SB(inode->i_sb)->s_daxdev;
 786
 787         if (unlikely(ext4_forced_shutdown(inode->i_sb)))
 788                 return -EIO;
 789
 790         /*
 791          * We don't support synchronous mappings for non-DAX files and
 792          * for DAX files if underneath dax_device is not synchronous.
 793          */
 794         if (!daxdev_mapping_supported(vma, dax_dev))
 795                 return -EOPNOTSUPP;
 796
 797         file_accessed(file);
 798         if (IS_DAX(file_inode(file))) {
 799                 vma->vm_ops = &ext4_dax_vm_ops;
 800                 vm_flags_set(vma, VM_HUGEPAGE);
 801         } else {
 802                 vma->vm_ops = &ext4_file_vm_ops;
 803         }
 804         return 0;
 805 }
 806
 807 static int ext4_sample_last_mounted(struct super_block *sb,
 808                                     struct vfsmount *mnt)
 809 {
 810         struct ext4_sb_info *sbi = EXT4_SB(sb);
 811         struct path path;
 812         char buf[64], *cp;
 813         handle_t *handle;
 814         int err;
 815
 816         if (likely(ext4_test_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED)))
 817                 return 0;
 818
 819         if (sb_rdonly(sb) || !sb_start_intwrite_trylock(sb))
 820                 return 0;
 821
 822         ext4_set_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED);
 823         /*
 824          * Sample where the filesystem has been mounted and
 825          * store it in the superblock for sysadmin convenience
 826          * when trying to sort through large numbers of block
 827          * devices or filesystem images.
 828          */
 829         memset(buf, 0, sizeof(buf));
 830         path.mnt = mnt;
 831         path.dentry = mnt->mnt_root;
 832         cp = d_path(&path, buf, sizeof(buf));
 833         err = 0;
 834         if (IS_ERR(cp))
 835                 goto out;
 836
 837         handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
 838         err = PTR_ERR(handle);
 839         if (IS_ERR(handle))
 840                 goto out;
 841         BUFFER_TRACE(sbi->s_sbh, "get_write_access");
 842         err = ext4_journal_get_write_access(handle, sb, sbi->s_sbh,
 843                                             EXT4_JTR_NONE);
 844         if (err)
 845                 goto out_journal;
 846         lock_buffer(sbi->s_sbh);
 847         strtomem_pad(sbi->s_es->s_last_mounted, cp, 0);
 848         ext4_superblock_csum_set(sb);
 849         unlock_buffer(sbi->s_sbh);
 850         ext4_handle_dirty_metadata(handle, NULL, sbi->s_sbh);
 851 out_journal:
 852         ext4_journal_stop(handle);
 853 out:
 854         sb_end_intwrite(sb);
 855         return err;
 856 }
 857
 858 static int ext4_file_open(struct inode *inode, struct file *filp)
 859 {
 860         int ret;
 861
 862         if (unlikely(ext4_forced_shutdown(inode->i_sb)))
 863                 return -EIO;
 864
 865         ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
 866         if (ret)
 867                 return ret;
 868
 869         ret = fscrypt_file_open(inode, filp);
 870         if (ret)
 871                 return ret;
 872
 873         ret = fsverity_file_open(inode, filp);
 874         if (ret)
 875                 return ret;
 876
 877         /*
 878          * Set up the jbd2_inode if we are opening the inode for
 879          * writing and the journal is present
 880          */
 881         if (filp->f_mode & FMODE_WRITE) {
 882                 ret = ext4_inode_attach_jinode(inode);
 883                 if (ret < 0)
 884                         return ret;
 885         }
 886
 887         filp->f_mode |= FMODE_NOWAIT | FMODE_CAN_ODIRECT;
 888         return dquot_file_open(inode, filp);
 889 }
 890
 891 /*
 892  * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
 893  * by calling generic_file_llseek_size() with the appropriate maxbytes
 894  * value for each.
 895  */
 896 loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
 897 {
 898         struct inode *inode = file->f_mapping->host;
 899         loff_t maxbytes;
 900
 901         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 902                 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
 903         else
 904                 maxbytes = inode->i_sb->s_maxbytes;
 905
 906         switch (whence) {
 907         default:
 908                 return generic_file_llseek_size(file, offset, whence,
 909                                                 maxbytes, i_size_read(inode));
 910         case SEEK_HOLE:
 911                 inode_lock_shared(inode);
 912                 offset = iomap_seek_hole(inode, offset,
 913                                          &ext4_iomap_report_ops);
 914                 inode_unlock_shared(inode);
 915                 break;
 916         case SEEK_DATA:
 917                 inode_lock_shared(inode);
 918                 offset = iomap_seek_data(inode, offset,
 919                                          &ext4_iomap_report_ops);
 920                 inode_unlock_shared(inode);
 921                 break;
 922         }
 923
 924         if (offset < 0)
 925                 return offset;
 926         return vfs_setpos(file, offset, maxbytes);
 927 }
 928
 929 const struct file_operations ext4_file_operations = {
 930         .llseek         = ext4_llseek,
 931         .read_iter      = ext4_file_read_iter,
 932         .write_iter     = ext4_file_write_iter,
 933         .iopoll         = iocb_bio_iopoll,
 934         .unlocked_ioctl = ext4_ioctl,
 935 #ifdef CONFIG_COMPAT
 936         .compat_ioctl   = ext4_compat_ioctl,
 937 #endif
 938         .mmap           = ext4_file_mmap,
 939         .open           = ext4_file_open,
 940         .release        = ext4_release_file,
 941         .fsync          = ext4_sync_file,
 942         .get_unmapped_area = thp_get_unmapped_area,
 943         .splice_read    = ext4_file_splice_read,
 944         .splice_write   = iter_file_splice_write,
 945         .fallocate      = ext4_fallocate,
 946         .fop_flags      = FOP_MMAP_SYNC | FOP_BUFFER_RASYNC |
 947                           FOP_DIO_PARALLEL_WRITE,
 948 };
 949
 950 const struct inode_operations ext4_file_inode_operations = {
 951         .setattr        = ext4_setattr,
 952         .getattr        = ext4_file_getattr,
 953         .listxattr      = ext4_listxattr,
 954         .get_inode_acl  = ext4_get_acl,
 955         .set_acl        = ext4_set_acl,
 956         .fiemap         = ext4_fiemap,
 957         .fileattr_get   = ext4_fileattr_get,
 958         .fileattr_set   = ext4_fileattr_set,
 959 };
 960