This patch set addresses problems hit when running liburing
500f9fbadef8-test.
- Patch 1 is a suggested fix to overloading of kiocb.private since it
can be written by iomap_dio_rw().
io_import_iovec() can fail a submission as follows:
/* buffer index only valid with fixed read/write, or buffer select */
if (req->rw.kiocb.private && !(req->flags & REQ_F_BUFFER_SELECT))
return -EINVAL;
so a read request fails with -EINVAL upon retry if iomap_dio_rw() has
written to iocb->private:
WRITE_ONCE(iocb->private, dio->submit.last_queue);
The suggested fix is use a separate variable to store buf_index.
- Patch 2 reverts c58c1f8343 which had changed the error for
REQ_NOWAIT requests to non-mq queue from -ENOTSUP to -EAGAIN.
/*
* Non-mq queues do not honor REQ_NOWAIT, so complete a bio
* with BLK_STS_AGAIN status in order to catch -EAGAIN and
* to give a chance to the caller to repeat request gracefully.
*/
if ((bio->bi_opf & REQ_NOWAIT) && !queue_is_mq(q)) {
status = BLK_STS_AGAIN;
goto end_io;
}
I'm not clear what the original reasoning was but io_wq_submit_work()
will call io_issue_sqe() continuously as long as -EAGAIN is returned.
I'm not sure if this could break something else.
I ran fio as specified in c58c1f8343 with this change and don't see any errors.
Bijan Mottahedeh (2):
io_uring: don't use kiocb.private to store buf_index
io_uring: mark REQ_NOWAIT for a non-mq queue as unspported
block/blk-core.c | 10 +++-------
fs/io_uring.c | 15 +++++++--------
2 files changed, 10 insertions(+), 15 deletions(-)
--
1.8.3.1
