"Theodore Tso" <tytso@xxxxxxx> writes: > Yeah, this change is going to problems. The basic idea here is if > when we request that a transaction to commit, will it issue a a > commit? If so, then fsync(2) doesn't need to issue a barrier (i.e., a > cache flush command). > > So for example, if a database does an overwriting write to a file > block which is already allocated, and then follows it up with a > fdatasync(2), there won't be any need to make any metadata changes as > part of writing out the changed block. Hence, we won't need to start > a new jbd2 transaction, and in that case, current transaction has > already commited, so the jbd2 layer won't need to do anything, and so > it won't have issued a commit. In that case, > jbd2_trans_will_send_data_barrier() needs to return false, so that > fdatasync(2) will actually issue a cache flush command. > > The patch you've proposed will cause fdatasync(2) to not issue a > barrier, which could lead to the write to the database file getting > lost after a power fail event, which would make the database > adminisrtator very sad. So because no metadata changed, jbd2 will not issue a barrier to end the transaction? How can we fix this then? Is there some way that jbd2 can know whether file data has been written, and thus, issue the barrier to close the transaction?
