When fio repeats same workload on zoned block devices, zbd_file_reset() is called before each repetition to reset target zones. If a thread runs such repeated workload with asynchronous IOs, the thread calls zbd_file_reset() and reset target zones even when the IOs are in-flight. This causes write pointer move during in-flight IOs and results in unexpected IO results. Especially if write requests are in-flight, they fail with unaligned write command error. A single thread may do both the zone reset and the write request submit, recursive zone locks can not prevent the zone reset during the writes in-flight. To avoid the zone reset during asynchronous IOs, add io_u_quiesce() call before the zone reset in zbd_file_reset(). This ensures that the in- flight IOs are completed before the zone reset. Signed-off-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@xxxxxxx> --- zbd.c | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/zbd.c b/zbd.c index d16b890f..5f0799b0 100644 --- a/zbd.c +++ b/zbd.c @@ -991,6 +991,13 @@ void zbd_file_reset(struct thread_data *td, struct fio_file *f) swd = zbd_process_swd(td, f, SET_SWD); dprint(FD_ZBD, "%s(%s): swd = %" PRIu64 "\n", __func__, f->file_name, swd); + + /* + * When the thread repeats workload on the same file, ensure the + * async IOs by previous repetition completed before zone reset. + */ + io_u_quiesce(td); + /* * If data verification is enabled reset the affected zones before * writing any data to avoid that a zone reset has to be issued while -- 2.29.2
