- stable@ 19.03.2021, 08:29, "Dmitry Monakhov" : > Otherwise we may endup blocking on pipe or socket. > > Fixes: 6d5d5ac ("io_uring: extend async work merge") > Testcase: https://github.com/dmonakhov/liburing/commit/16d171b6ef9d68e6db66650a83d98c5c721d01f6 > Signed-off-by: Dmitry Monakhov > --- >  fs/io_uring.c | 3 +++ >  1 file changed, 3 insertions(+) > > diff --git a/fs/io_uring.c b/fs/io_uring.c > index 478df7e..848657c 100644 > --- a/fs/io_uring.c > +++ b/fs/io_uring.c > @@ -2183,6 +2183,9 @@ static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, >  static struct async_list *io_async_list_from_req(struct io_ring_ctx *ctx, >                                                   struct io_kiocb *req) >  { > + if (!(req->flags & REQ_F_ISREG)) > + return NULL; > + IMHO it is reasonable to completely disable io_should_merge logic because even with the this fix it still affected by latency spikes like follows: ->submit_read: req1[slow_hdd, sector=z] ->submit_read: req2[nvme, sector=X] -> wait(req2) -> fast ->submit_read: req3[nvme, sector=Y] --> wait(req3) ->slow if completes if X and Y belongs to same page merge logic will wait for req1 to complete