fs/ext4/file.c

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