If we are doing locked DIO, the OSC & LDLM locks are
released at the end of cl_io_loop, ie, before we wait for
parallel DIO at the llite layer.
This is problematic because the locks are released before
i/o done using them is complete; this can lead to data
inconsistencies. (And at least one LBUG, see LU-14805.)
The easiest solution for now is only do parallel DIO when
working lockless (which is the default; DIO only switches
to locked to manage conflicts with buffered i/o).
This problem & fix apply to AIO as well as parallel DIO.
Signed-off-by: Patrick Farrell <pfarrell@whamcloud.com>
Change-Id: If98a0551d6dde54220b406b26e978e284a6b1ebf
Reviewed-on: https://review.whamcloud.com/44131
Tested-by: jenkins <devops@whamcloud.com>
Reviewed-by: Wang Shilong <wshilong@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>
else
vio->u.readwrite.vui_read += tot_bytes;
else
vio->u.readwrite.vui_read += tot_bytes;
- /* If async dio submission is not allowed, we must wait here. */
- if (is_sync_kiocb(iocb) && !io->ci_parallel_dio) {
+ /* We cannot do async submission - for AIO or regular DIO - unless
+ * lockless because it causes us to release the lock early.
+ *
+ * There are also several circumstances in which we must disable
+ * parallel DIO, so we check if it is enabled.
+ *
+ * The check for "is_sync_kiocb" excludes AIO, which does not need to
+ * be disabled in these situations.
+ */
+ if (io->ci_dio_lock || (is_sync_kiocb(iocb) && !io->ci_parallel_dio)) {
+ /* Wait here rather than doing async submission */
rc2 = cl_sync_io_wait_recycle(env, &aio->cda_sync, 0, 0);
if (result == 0 && rc2)
result = rc2;
rc2 = cl_sync_io_wait_recycle(env, &aio->cda_sync, 0, 0);
if (result == 0 && rc2)
result = rc2;