io_uring: improve task work cache utilization
While profiling task_work intensive workloads, I noticed that most of
the time in tctx_task_work() is spending stalled on loading 'req'. This
is one of the unfortunate side effects of using linked lists,
particularly when they end up being passe around.
Prefetch the next request, if there is one. There's a sufficient amount
of work in between that this makes it available for the next loop.
While fiddling with the cache layout, move the link outside of the
hot completion cacheline. It's rarely used in hot workloads, so better
to bring in kbuf which is used for networked loads with provided buffers.
This reduces tctx_task_work() overhead from ~3% to 1-1.5% in my testing.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
diff --git a/fs/io_uring.c b/fs/io_uring.c
index a76e91f..bb40c80 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -928,7 +928,6 @@ struct io_kiocb {
struct io_wq_work_node comp_list;
atomic_t refs;
atomic_t poll_refs;
- struct io_kiocb *link;
struct io_task_work io_task_work;
/* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
struct hlist_node hash_node;
@@ -939,6 +938,7 @@ struct io_kiocb {
/* custom credentials, valid IFF REQ_F_CREDS is set */
/* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
struct io_buffer *kbuf;
+ struct io_kiocb *link;
const struct cred *creds;
struct io_wq_work work;
};
@@ -2451,6 +2451,8 @@ static void handle_prev_tw_list(struct io_wq_work_node *node,
struct io_kiocb *req = container_of(node, struct io_kiocb,
io_task_work.node);
+ prefetch(container_of(next, struct io_kiocb, io_task_work.node));
+
if (req->ctx != *ctx) {
if (unlikely(!*uring_locked && *ctx))
ctx_commit_and_unlock(*ctx);
@@ -2483,6 +2485,8 @@ static void handle_tw_list(struct io_wq_work_node *node,
struct io_kiocb *req = container_of(node, struct io_kiocb,
io_task_work.node);
+ prefetch(container_of(next, struct io_kiocb, io_task_work.node));
+
if (req->ctx != *ctx) {
ctx_flush_and_put(*ctx, locked);
*ctx = req->ctx;
