On 11/30/18 10:12 AM, Bart Van Assche wrote: > On Fri, 2018-11-30 at 09:56 -0700, Jens Axboe wrote: >> We can't wait for polled events to complete, as they may require active >> polling from whoever submitted it. If that is the same task that is >> submitting new IO, we could deadlock waiting for IO to complete that >> this task is supposed to be completing itself. >> >> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx> >> --- >> fs/block_dev.c | 10 +++++++++- >> 1 file changed, 9 insertions(+), 1 deletion(-) >> >> diff --git a/fs/block_dev.c b/fs/block_dev.c >> index 6de8d35f6e41..ebc3d5a0f424 100644 >> --- a/fs/block_dev.c >> +++ b/fs/block_dev.c >> @@ -402,8 +402,16 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages) >> >> nr_pages = iov_iter_npages(iter, BIO_MAX_PAGES); >> if (!nr_pages) { >> - if (iocb->ki_flags & IOCB_HIPRI) >> + if (iocb->ki_flags & IOCB_HIPRI) { >> bio->bi_opf |= REQ_HIPRI; >> + /* >> + * For async polled IO, we can't wait for >> + * requests to complete, as they may also be >> + * polled and require active reaping. >> + */ >> + if (!is_sync) >> + bio->bi_opf |= REQ_NOWAIT; >> + } >> >> qc = submit_bio(bio); >> WRITE_ONCE(iocb->ki_cookie, qc); > > Setting REQ_NOWAIT from inside the block layer will make the code that > submits requests harder to review. Have you considered to make this code > fail I/O if REQ_NOWAIT has not been set and to require that the context > that submits I/O sets REQ_NOWAIT? It's technically still feasible to do for sync polled IO, it's only the async case that makes it a potential deadlock. -- Jens Axboe