]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
ext4: Fix data integrity writeout issues in nojournal mode
authorJan Kara <jack@suse.cz>
Mon, 27 Jul 2026 10:49:35 +0000 (12:49 +0200)
committerChristian Brauner <brauner@kernel.org>
Mon, 27 Jul 2026 14:25:33 +0000 (16:25 +0200)
Several racing fsyncs on ext4 in nojournal mode could result in some
fsync returning earlier than all metadata buffers were properly
persisted. Also ext4_fsync() in nojournal mode was somewhat inefficient
because it was always writing out the inode regardless whether it was
dirty or not.

Fix these issues by using new .sync_inode_metadata method which makes
sure all inode related metadata is written to disk during any
WB_SYNC_ALL writeback in nojournal mode. This also somewhat simplifies
the nojournal mode fsync handling.

Signed-off-by: Jan Kara <jack@suse.cz>
Link: https://patch.msgid.link/20260727104923.3828017-37-jack@suse.cz
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
fs/ext4/ext4.h
fs/ext4/fsync.c
fs/ext4/inode.c
fs/ext4/super.c

index 64f8f63f4415e7ac9c712856c57059dcc1cdda01..0f06155a35a694a057e3ca546a821fd819089c37 100644 (file)
@@ -3166,6 +3166,7 @@ extern struct inode *__ext4_iget(struct super_block *sb, unsigned long ino,
        __ext4_iget((sb), (ino), (flags), __func__, __LINE__)
 
 extern int  ext4_write_inode(struct inode *, struct writeback_control *);
+extern int  ext4_sync_inode_metadata(struct inode *, struct writeback_control *);
 extern int  ext4_setattr(struct mnt_idmap *, struct dentry *,
                         struct iattr *);
 extern u32  ext4_dio_alignment(struct inode *inode);
index b7ea4433f4bebefbd544d598fb8f851096a213f3..2999c2cc8fcfa0f68edc08b3878f6ed5edcfbc7b 100644 (file)
@@ -46,7 +46,6 @@
 static int ext4_sync_parent(struct inode *inode)
 {
        struct dentry *dentry, *next;
-       struct mapping_metadata_bhs *mmb;
        int ret = 0;
 
        if (!ext4_test_inode_state(inode, EXT4_STATE_NEWENTRY))
@@ -69,12 +68,6 @@ static int ext4_sync_parent(struct inode *inode)
                 * through ext4_evict_inode()) and so we are safe to flush
                 * metadata blocks and the inode.
                 */
-               mmb = ext4_i_metadata_bhs(inode);
-               if (mmb) {
-                       ret = mmb_sync(mmb);
-                       if (ret)
-                               break;
-               }
                ret = sync_inode_metadata(inode, 1);
                if (ret)
                        break;
@@ -87,22 +80,11 @@ static int ext4_fsync_nojournal(struct file *file, loff_t start, loff_t end,
                                int datasync, bool *needs_barrier)
 {
        struct inode *inode = file->f_inode;
-       struct writeback_control wbc = {
-               .sync_mode = WB_SYNC_ALL,
-               .nr_to_write = 0,
-       };
        int ret;
 
-       ret = mmb_fsync_noflush(file, ext4_i_metadata_bhs(inode),
-                               start, end, datasync);
+       ret = sync_inode_metadata(inode, 1);
        if (ret)
                return ret;
-
-       /* Force writeout of inode table buffer to disk */
-       ret = ext4_write_inode(inode, &wbc);
-       if (ret)
-               return ret;
-
        ret = ext4_sync_parent(inode);
 
        if (test_opt(inode->i_sb, BARRIER))
@@ -160,6 +142,10 @@ int ext4_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
        if (sb_rdonly(inode->i_sb))
                goto out;
 
+       ret = file_write_and_wait_range(file, start, end);
+       if (ret)
+               goto out;
+
        if (!EXT4_SB(inode->i_sb)->s_journal) {
                ret = ext4_fsync_nojournal(file, start, end, datasync,
                                           &needs_barrier);
@@ -168,10 +154,6 @@ int ext4_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
                goto out;
        }
 
-       ret = file_write_and_wait_range(file, start, end);
-       if (ret)
-               goto out;
-
        /*
         *  The caller's filemap_fdatawrite()/wait will sync the data.
         *  Metadata is in the journal, we wait for proper transaction to
index e6acef486ee1376f616e4441ff5f173364f823d2..7a1f961cd11cebf4c3a45ae449fcae5051071395 100644 (file)
@@ -5799,6 +5799,10 @@ out_brelse:
  * ext4_mark_inode_dirty().  This is a correctness thing for WB_SYNC_ALL
  * writeback.
  *
+ * For nojournal mode all the work is done in ext4_sync_inode_metadata()
+ * because inode content is already copied into raw inode buffer and inode
+ * is marked with I_METADATA_WRITEBACK.
+ *
  * Note that we are absolutely dependent upon all inode dirtiers doing the
  * right thing: they *must* call mark_inode_dirty() after dirtying info in
  * which we are interested.
@@ -5824,42 +5828,54 @@ int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
        if (unlikely(err))
                return err;
 
-       if (EXT4_SB(inode->i_sb)->s_journal) {
-               if (ext4_journal_current_handle()) {
-                       ext4_debug("called recursively, non-PF_MEMALLOC!\n");
-                       dump_stack();
-                       return -EIO;
-               }
+       if (!EXT4_SB(inode->i_sb)->s_journal)
+               return 0;
 
-               /*
-                * No need to force transaction in WB_SYNC_NONE mode. Also
-                * ext4_sync_fs() will force the commit after everything is
-                * written.
-                */
-               if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
-                       return 0;
+       if (ext4_journal_current_handle()) {
+               ext4_debug("called recursively, non-PF_MEMALLOC!\n");
+               dump_stack();
+               return -EIO;
+       }
+
+       /*
+        * No need to force transaction in WB_SYNC_NONE mode. Also
+        * ext4_sync_fs() will force the commit after everything is
+        * written.
+        */
+       if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
+               return 0;
 
-               err = ext4_fc_commit(EXT4_SB(inode->i_sb)->s_journal,
+       return ext4_fc_commit(EXT4_SB(inode->i_sb)->s_journal,
                                                EXT4_I(inode)->i_sync_tid);
-       } else {
-               struct ext4_iloc iloc;
+}
+
+int ext4_sync_inode_metadata(struct inode *inode, struct writeback_control *wbc)
+{
+       struct ext4_iloc iloc;
+       struct mapping_metadata_bhs *mmb;
+       int err;
 
-               err = __ext4_get_inode_loc_noinmem(inode, &iloc);
+       /* We should only get here in nojournal mode */
+       if (WARN_ON_ONCE(EXT4_SB(inode->i_sb)->s_journal))
+               return -EFSCORRUPTED;
+
+       err = __ext4_get_inode_loc_noinmem(inode, &iloc);
+       if (err)
+               return err;
+       mmb = READ_ONCE(EXT4_I(inode)->i_metadata_bhs);
+       if (mmb) {
+               err = mmb_sync(mmb);
                if (err)
-                       return err;
-               /*
-                * sync(2) will flush the whole buffer cache. No need to do
-                * it here separately for each inode.
-                */
-               if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
-                       sync_dirty_buffer(iloc.bh);
-               if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
-                       ext4_error_inode_block(inode, iloc.bh->b_blocknr, EIO,
-                                              "IO error syncing inode");
-                       err = -EIO;
-               }
-               brelse(iloc.bh);
+                       goto out;
        }
+       sync_dirty_buffer(iloc.bh);
+       if (buffer_write_io_error(iloc.bh)) {
+               ext4_error_inode_block(inode, iloc.bh->b_blocknr, EIO,
+                                      "IO error syncing inode");
+               err = -EIO;
+       }
+out:
+       brelse(iloc.bh);
        return err;
 }
 
@@ -6407,6 +6423,20 @@ int ext4_mark_iloc_dirty(handle_t *handle,
        /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
        err = ext4_do_update_inode(handle, inode, iloc);
        put_bh(iloc->bh);
+       /*
+        * Mark that there's metadata writeout pending for the inode so that it
+        * gets properly flushed on fsync(2) and similar.
+        */
+       if (!EXT4_SB(inode->i_sb)->s_journal) {
+               /*
+                * Inode didn't need to go through dirtying, make sure it is
+                * attached to wb so that writeback can handle it.
+                */
+               spin_lock(&inode->i_lock);
+               inode_attach_wb(inode, NULL);
+               spin_unlock(&inode->i_lock);
+               set_inode_metadata_writeback(inode);
+       }
        return err;
 }
 
index 8671fa1209dd5e92bb9a4dd8310614bb1d2e4ef4..ae33f5bcb133140f9d019faee2386f0584d17409 100644 (file)
@@ -1608,9 +1608,13 @@ static int ext4_nfs_commit_metadata(struct inode *inode)
        struct writeback_control wbc = {
                .sync_mode = WB_SYNC_ALL
        };
+       int ret;
 
        trace_ext4_nfs_commit_metadata(inode);
-       return ext4_write_inode(inode, &wbc);
+       ret = ext4_write_inode(inode, &wbc);
+       if (!ret && inode_state_read_once(inode) & I_METADATA_WRITEBACK)
+               ret = ext4_sync_inode_metadata(inode, &wbc);
+       return ret;
 }
 
 #ifdef CONFIG_QUOTA
@@ -1667,6 +1671,7 @@ static const struct super_operations ext4_sops = {
        .free_inode     = ext4_free_in_core_inode,
        .destroy_inode  = ext4_destroy_inode,
        .write_inode    = ext4_write_inode,
+       .sync_inode_metadata = ext4_sync_inode_metadata,
        .dirty_inode    = ext4_dirty_inode,
        .drop_inode     = ext4_drop_inode,
        .evict_inode    = ext4_evict_inode,