On Fri 13-10-23 10:33:08, Dave Chinner wrote: > On Thu, Oct 12, 2023 at 10:59:37AM +0200, Jan Kara wrote: > > On Thu 12-10-23 11:26:15, Dave Chinner wrote: > > > On Wed, Oct 11, 2023 at 04:21:55PM +0200, Jan Kara wrote: > > > > Gao Xiang has reported that on ext4 O_SYNC direct IO does not properly > > > > sync file size update and thus if we crash at unfortunate moment, the > > > > file can have smaller size although O_SYNC IO has reported successful > > > > completion. The problem happens because update of on-disk inode size is > > > > handled in ext4_dio_write_iter() *after* iomap_dio_rw() (and thus > > > > dio_complete() in particular) has returned and generic_file_sync() gets > > > > called by dio_complete(). Fix the problem by handling on-disk inode size > > > > update directly in our ->end_io completion handler. > > > > > > > > References: https://lore.kernel.org/all/02d18236-26ef-09b0-90ad-030c4fe3ee20@xxxxxxxxxxxxxxxxx > > > > Reported-by: Gao Xiang <hsiangkao@xxxxxxxxxxxxxxxxx> > > > > Signed-off-by: Jan Kara <jack@xxxxxxx> > > > > --- > > > > fs/ext4/file.c | 139 ++++++++++++++++++------------------------------- > > > > 1 file changed, 52 insertions(+), 87 deletions(-) > > > ..... > > > > @@ -388,9 +342,28 @@ static ssize_t ext4_handle_inode_extension(struct inode *inode, loff_t offset, > > > > */ > > > > if (inode->i_nlink) > > > > ext4_orphan_del(NULL, inode); > > > > + return; > > > > } > > > > + /* > > > > + * If i_disksize got extended due to writeback of delalloc blocks while > > > > + * the DIO was running we could fail to cleanup the orphan list in > > > > + * ext4_handle_inode_extension(). Do it now. > > > > + */ > > > > + if (!list_empty(&EXT4_I(inode)->i_orphan) && inode->i_nlink) { > > > > + handle_t *handle = ext4_journal_start(inode, EXT4_HT_INODE, 2); > > > > > > So this has to be called after the DIO write completes and calls > > > ext4_handle_inode_extension()? > > > > Yes, if the write was setup as extending one ('extend' is set to true in > > ext4_dio_write_iter()). > > Then that is worth a comment to document the constraint for anyone > that is trying to understand how ext4 is using the iomap DIO code. Fair enough, comment added. > > > > @@ -606,9 +570,8 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from) > > > > dio_flags, NULL, 0); > > > > if (ret == -ENOTBLK) > > > > ret = 0; > > > > - > > > > if (extend) > > > > - ret = ext4_handle_inode_extension(inode, offset, ret, count); > > > > + ext4_inode_extension_cleanup(inode, ret); > > > > > > Because this doesn't wait for AIO DIO to complete and actually > > > extend the file before running the cleanup code... > > > > As Gao wrote, ext4 sets IOMAP_DIO_FORCE_WAIT if 'extend' is set (see > > ext4_dio_write_checks()) so if we get to calling > > ext4_inode_extension_cleanup() we are guaranteed the IO has already > > completed. > > Ugh. That's a pretty nasty undocumented landmine. It definitely > needs a comment (or better, a WARN_ON_ONCE()) to document that this > code -only- works if AIO is disabled. This isn't for ext4 > developers, it's for people working on the iomap code to understand > that ext4 has some really non-obvious constraints in it's DIO code > paths and that's why the landmine is not being stepped on.... OK. Ext4 has always been this way so I never felt the need to document it but you're right. I've added: if (extend) { /* * We always perform extending DIO write synchronously so by * now the IO is completed and ext4_handle_inode_extension() * was called. Cleanup the inode in case of error or race with * writeback of delalloc blocks. */ WARN_ON_ONCE(ret == -EIOCBQUEUED); ext4_inode_extension_cleanup(inode, ret); } Thanks for the suggestions! Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
