Running fio polled tests with a large number of jobs ends up with stuck polling threads. I think this because of polling for requests that have completed with -EAGAIN. Running with the RFC applied, and with sufficiently large nvme timeout values, the fio tests complete. The RFC creates a retry list similar to the done list. I'm not sure if that's the best approach and whether there may be ordering issues processing the two lists but I haven't seen any problems. Bijan Mottahedeh (1): io_uring: process requests completed with -EAGAIN on poll list fs/io_uring.c | 29 ++++++++++++++++++++++++++--- 1 file changed, 26 insertions(+), 3 deletions(-) -- 1.8.3.1
