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