t/io_uring: enable SQ thread poll mode
With this, we can do IO without ever entering the kernel.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
diff --git a/t/io_uring.c b/t/io_uring.c
index 1c0d153..58a1c45 100644
--- a/t/io_uring.c
+++ b/t/io_uring.c
@@ -33,6 +33,7 @@
unsigned *tail;
unsigned *ring_mask;
unsigned *ring_entries;
+ unsigned *flags;
unsigned *array;
};
@@ -250,6 +251,7 @@
static void *submitter_fn(void *data)
{
struct submitter *s = data;
+ struct io_sq_ring *ring = &s->sq_ring;
int ret, prepped;
printf("submitter=%d\n", gettid());
@@ -273,13 +275,30 @@
else
to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
- ret = io_uring_enter(s, to_submit, to_wait,
- IORING_ENTER_GETEVENTS);
- s->calls++;
+ /*
+ * Only need to call io_uring_enter if we're not using SQ thread
+ * poll, or if IORING_SQ_NEED_WAKEUP is set.
+ */
+ if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
+ ret = io_uring_enter(s, to_submit, to_wait,
+ IORING_ENTER_GETEVENTS);
+ s->calls++;
+ }
- this_reap = reap_events(s);
- if (this_reap == -1)
- break;
+ /*
+ * For non SQ thread poll, we already got the events we needed
+ * through the io_uring_enter() above. For SQ thread poll, we
+ * need to loop here until we find enough events.
+ */
+ this_reap = 0;
+ do {
+ int r;
+ r = reap_events(s);
+ if (r == -1)
+ break;
+ else if (r > 0)
+ this_reap += r;
+ } while (sq_thread_poll && this_reap < to_wait);
s->reaps += this_reap;
if (ret >= 0) {
@@ -374,6 +393,7 @@
sring->tail = ptr + p.sq_off.tail;
sring->ring_mask = ptr + p.sq_off.ring_mask;
sring->ring_entries = ptr + p.sq_off.ring_entries;
+ sring->flags = ptr + p.sq_off.flags;
sring->array = ptr + p.sq_off.array;
sq_ring_mask = *sring->ring_mask;
