io_uring: allow defer completion for aux posted cqes

Multishot ops cannot use the compl_reqs list as the request must stay in
the poll list, but that means they need to run each completion without
benefiting from batching.

Here introduce batching infrastructure for only small (ie 16 byte)
CQEs. This restriction is ok because there are no use cases posting 32
byte CQEs.

In the ring keep a batch of up to 16 posted results, and flush in the same
way as compl_reqs.

16 was chosen through experimentation on a microbenchmark ([1]), as well
as trying not to increase the size of the ring too much. This increases
the size to 1472 bytes from 1216.

[1]: https://github.com/DylanZA/liburing/commit/9ac66b36bcf4477bfafeff1c5f107896b7ae31cf
Run with $ make -j && ./benchmark/reg.b -s 1 -t 2000 -r 10
Gives results:
baseline	8309 k/s
8		18807 k/s
16		19338 k/s
32		20134 k/s

Suggested-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Dylan Yudaken <dylany@meta.com>
Link: https://lore.kernel.org/r/20221124093559.3780686-5-dylany@meta.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

1 files changed, 24 insertions, 3 deletions

diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index 4888fe834920..28635e3e578a 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -167,7 +167,8 @@ EXPORT_SYMBOL(io_uring_get_socket);
 
 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
 {
-	if (!wq_list_empty(&ctx->submit_state.compl_reqs))
+	if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
+	    ctx->submit_state.cqes_count)
 		__io_submit_flush_completions(ctx);
 }
 
@@ -802,6 +803,21 @@ bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags
 	return false;
 }
 
+static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
+	__must_hold(&ctx->uring_lock)
+{
+	struct io_submit_state *state = &ctx->submit_state;
+	unsigned int i;
+
+	lockdep_assert_held(&ctx->uring_lock);
+	for (i = 0; i < state->cqes_count; i++) {
+		struct io_uring_cqe *cqe = &state->cqes[i];
+
+		io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags, true);
+	}
+	state->cqes_count = 0;
+}
+
 bool io_post_aux_cqe(struct io_ring_ctx *ctx,
 		     u64 user_data, s32 res, u32 cflags,
 		     bool allow_overflow)
@@ -1323,6 +1339,9 @@ static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
 	struct io_submit_state *state = &ctx->submit_state;
 
 	io_cq_lock(ctx);
+	/* must come first to preserve CQE ordering in failure cases */
+	if (state->cqes_count)
+		__io_flush_post_cqes(ctx);
 	wq_list_for_each(node, prev, &state->compl_reqs) {
 		struct io_kiocb *req = container_of(node, struct io_kiocb,
 					    comp_list);
@@ -1332,8 +1351,10 @@ static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
 	}
 	io_cq_unlock_post(ctx);
 
-	io_free_batch_list(ctx, state->compl_reqs.first);
-	INIT_WQ_LIST(&state->compl_reqs);
+	if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
+		io_free_batch_list(ctx, state->compl_reqs.first);
+		INIT_WQ_LIST(&state->compl_reqs);
+	}
 }
 
 /*