[PATCH v2 7/8] io_uring: persistent req bulk allocation cache

[Pavel Begunkov <asml.silence@xxxxxxxxx>]

Save bulk allocated requests across io_uring_enter(), so lower QD also
could benefit from that. This is not much of an optimisation, and for
current cache sizes would probably affect only offloaded ~QD=1.

Signed-off-by: Pavel Begunkov <asml.silence@xxxxxxxxx>
---
 fs/io_uring.c | 38 ++++++++++++++++++++++++++------------
 1 file changed, 26 insertions(+), 12 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 5022eb4cb9a4..82df6171baae 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -834,6 +834,25 @@ static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 	return NULL;
 }
 
+static void io_init_submit_state(struct io_ring_ctx *ctx)
+{
+	struct io_submit_state *state = &ctx->submit_state;
+
+	state->mm = (ctx->flags & IORING_SETUP_SQPOLL) ? NULL : ctx->sqo_mm;
+
+	state->free_reqs = 0;
+	state->cur_req = 0;
+}
+
+static void io_clear_submit_state(struct io_ring_ctx *ctx)
+{
+	struct io_submit_state *state = &ctx->submit_state;
+
+	if (state->free_reqs)
+		kmem_cache_free_bulk(req_cachep, state->free_reqs,
+					&state->reqs[state->cur_req]);
+}
+
 static inline bool __req_need_defer(struct io_kiocb *req)
 {
 	struct io_ring_ctx *ctx = req->ctx;
@@ -1132,10 +1151,9 @@ static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx)
 	struct io_submit_state *state = &ctx->submit_state;
 
 	if (!state->free_reqs) {
-		size_t sz;
+		size_t sz = ARRAY_SIZE(state->reqs);
 		int ret;
 
-		sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
 		ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
 
 		/*
@@ -4759,9 +4777,6 @@ static void io_submit_end(struct io_ring_ctx *ctx)
 	struct io_submit_state *state = &ctx->submit_state;
 
 	io_file_put(state);
-	if (state->free_reqs)
-		kmem_cache_free_bulk(req_cachep, state->free_reqs,
-					&state->reqs[state->cur_req]);
 	if (state->link)
 		io_queue_link_head(state->link);
 }
@@ -4774,7 +4789,6 @@ static void io_submit_start(struct io_ring_ctx *ctx, unsigned int max_ios,
 {
 	struct io_submit_state *state = &ctx->submit_state;
 
-	state->free_reqs = 0;
 	state->file = NULL;
 	state->ios_left = max_ios;
 
@@ -5765,12 +5779,6 @@ static int io_sq_offload_start(struct io_ring_ctx *ctx,
 	int ret;
 
 	init_waitqueue_head(&ctx->sqo_wait);
-	mmgrab(current->mm);
-	ctx->sqo_mm = current->mm;
-
-	ctx->submit_state.mm = NULL;
-	if (!(ctx->flags & IORING_SETUP_SQPOLL))
-		ctx->submit_state.mm = ctx->sqo_mm;
 
 	if (ctx->flags & IORING_SETUP_SQPOLL) {
 		ret = -EPERM;
@@ -6146,6 +6154,8 @@ static void io_ring_ctx_free(struct io_ring_ctx *ctx)
 	if (ctx->sqo_mm)
 		mmdrop(ctx->sqo_mm);
 
+	io_clear_submit_state(ctx);
+
 	io_iopoll_reap_events(ctx);
 	io_sqe_buffer_unregister(ctx);
 	io_sqe_files_unregister(ctx);
@@ -6584,6 +6594,10 @@ static int io_uring_create(unsigned entries, struct io_uring_params *p)
 	if (ret)
 		goto err;
 
+	mmgrab(current->mm);
+	ctx->sqo_mm = current->mm;
+	io_init_submit_state(ctx);
+
 	ret = io_sq_offload_start(ctx, p);
 	if (ret)
 		goto err;
-- 
2.24.0