| Message ID | 20190801234031.29561-18-mehta.aaru20@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Add support for io_uring | expand |
On Fri, Aug 02, 2019 at 05:10:31AM +0530, Aarushi Mehta wrote: > Signed-off-by: Aarushi Mehta <mehta.aaru20@gmail.com> > --- > block/io_uring.c | 17 ++++++++++++++++- > 1 file changed, 16 insertions(+), 1 deletion(-) How does SQPOLL performance compare? > diff --git a/block/io_uring.c b/block/io_uring.c > index 1553cd2e58..2a1d79704a 100644 > --- a/block/io_uring.c > +++ b/block/io_uring.c > @@ -288,6 +288,17 @@ static int ioq_submit(LuringState *s) > *sqes = luringcb->sqeq; > QSIMPLEQ_REMOVE_HEAD(&s->io_q.submit_queue, next); > } > + /* > + * io_uring_submit() returns sqes in ring for kernel side > + * submission polling and sets wakeup flag if needed. > + * > + * It is not possible for any sqes to have already been > + * submitted by the sq_poll as the writes are only made visible > + * to the kernel in this function. > + * > + * For normal I/O, it returns the actual submitted requests > + * from io_uring_enter() > + */ > ret = io_uring_submit(&s->ring); > trace_luring_io_uring_submit(s, ret); > /* Prevent infinite loop if submission is refused */ > @@ -525,7 +536,11 @@ LuringState *luring_init(Error **errp) > s = g_new0(LuringState, 1); > trace_luring_init_state(s, sizeof(*s)); > struct io_uring *ring = &s->ring; > - rc = io_uring_queue_init(MAX_EVENTS, ring, 0); > + > + rc = io_uring_queue_init(MAX_EVENTS, ring, IORING_SETUP_SQPOLL); > + if (rc == -EOPNOTSUPP) { > + rc = io_uring_queue_init(MAX_EVENTS, ring, 0); Please use 4-space indentation.
On Fri, Aug 2, 2019 at 12:50 AM Aarushi Mehta <mehta.aaru20@gmail.com> wrote: > + rc = io_uring_queue_init(MAX_EVENTS, ring, IORING_SETUP_SQPOLL); > + if (rc == -EOPNOTSUPP) { > + rc = io_uring_queue_init(MAX_EVENTS, ring, 0); > + } IORING_SETUP_SQPOLL is only allowed when the user has CAP_SYS_ADMIN (e.g. they are root). Otherwise it fails with -EPERM. This patch breaks non-root QEMU usage. Please handle -EPERM like -EOPNOTSUPP so it falls back to non-polling mode. Stefan
diff --git a/block/io_uring.c b/block/io_uring.c index 1553cd2e58..2a1d79704a 100644 --- a/block/io_uring.c +++ b/block/io_uring.c @@ -288,6 +288,17 @@ static int ioq_submit(LuringState *s) *sqes = luringcb->sqeq; QSIMPLEQ_REMOVE_HEAD(&s->io_q.submit_queue, next); } + /* + * io_uring_submit() returns sqes in ring for kernel side + * submission polling and sets wakeup flag if needed. + * + * It is not possible for any sqes to have already been + * submitted by the sq_poll as the writes are only made visible + * to the kernel in this function. + * + * For normal I/O, it returns the actual submitted requests + * from io_uring_enter() + */ ret = io_uring_submit(&s->ring); trace_luring_io_uring_submit(s, ret); /* Prevent infinite loop if submission is refused */ @@ -525,7 +536,11 @@ LuringState *luring_init(Error **errp) s = g_new0(LuringState, 1); trace_luring_init_state(s, sizeof(*s)); struct io_uring *ring = &s->ring; - rc = io_uring_queue_init(MAX_EVENTS, ring, 0); + + rc = io_uring_queue_init(MAX_EVENTS, ring, IORING_SETUP_SQPOLL); + if (rc == -EOPNOTSUPP) { + rc = io_uring_queue_init(MAX_EVENTS, ring, 0); + } if (rc < 0) { error_setg_errno(errp, errno, "failed to init linux io_uring ring"); g_free(s);
Signed-off-by: Aarushi Mehta <mehta.aaru20@gmail.com> --- block/io_uring.c | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-)