[PATCH v3 2/6] io_uring: place io_submit_state into ctx

[Pavel Begunkov <asml.silence@gmail.com>]

io_submit_state is used only during submmission and holding
ctx->uring_lock, so only one instance is used at a time. Move it into
struct io_ring_ctx, so it:
- doesn't consume on-stack memory
- persists across io_uring_enter
- available without passing it through the call-stack

The last point is very useful to make opcode handlers manage their
resources themselfs, like splice would. Also, it's a base for other
hackish optimisations in the future.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
 fs/io_uring.c | 75 +++++++++++++++++++++++++++------------------------
 1 file changed, 40 insertions(+), 35 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 6f3998e6475a..6109969709ff 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -197,6 +197,27 @@ struct fixed_file_data {
 	struct completion		done;
 };
 
+#define IO_PLUG_THRESHOLD		2
+#define IO_IOPOLL_BATCH			8
+
+struct io_submit_state {
+	/*
+	 * io_kiocb alloc cache
+	 */
+	void			*reqs[IO_IOPOLL_BATCH];
+	unsigned int		free_reqs;
+	unsigned int		cur_req;
+
+	/*
+	 * File reference cache
+	 */
+	struct file		*file;
+	unsigned int		fd;
+	unsigned int		has_refs;
+	unsigned int		used_refs;
+	unsigned int		ios_left;
+};
+
 struct io_ring_ctx {
 	struct {
 		struct percpu_ref	refs;
@@ -310,6 +331,9 @@ struct io_ring_ctx {
 		spinlock_t		inflight_lock;
 		struct list_head	inflight_list;
 	} ____cacheline_aligned_in_smp;
+
+	/* protected by uring_lock */
+	struct io_submit_state		submit_state;
 };
 
 /*
@@ -575,27 +599,6 @@ struct io_kiocb {
 	struct io_wq_work	work;
 };
 
-#define IO_PLUG_THRESHOLD		2
-#define IO_IOPOLL_BATCH			8
-
-struct io_submit_state {
-	/*
-	 * io_kiocb alloc cache
-	 */
-	void			*reqs[IO_IOPOLL_BATCH];
-	unsigned		int free_reqs;
-	unsigned		int cur_req;
-
-	/*
-	 * File reference cache
-	 */
-	struct file		*file;
-	unsigned int		fd;
-	unsigned int		has_refs;
-	unsigned int		used_refs;
-	unsigned int		ios_left;
-};
-
 struct io_op_def {
 	/* needs req->io allocated for deferral/async */
 	unsigned		async_ctx : 1;
@@ -1158,11 +1161,11 @@ static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
 	return NULL;
 }
 
-static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
-				   struct io_submit_state *state)
+static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx)
 {
 	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
 	struct io_kiocb *req;
+	struct io_submit_state *state = &ctx->submit_state;
 
 	if (!state->free_reqs) {
 		size_t sz;
@@ -4475,10 +4478,10 @@ static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
 	return table->files[index & IORING_FILE_TABLE_MASK];;
 }
 
-static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
-			   const struct io_uring_sqe *sqe)
+static int io_req_set_file(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 {
 	struct io_ring_ctx *ctx = req->ctx;
+	struct io_submit_state *state = &ctx->submit_state;
 	unsigned flags;
 	int fd;
 
@@ -4717,7 +4720,7 @@ static inline void io_queue_link_head(struct io_kiocb *req)
 				IOSQE_IO_HARDLINK | IOSQE_ASYNC)
 
 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
-			  struct io_submit_state *state, struct io_kiocb **link)
+			  struct io_kiocb **link)
 {
 	const struct cred *old_creds = NULL;
 	struct io_ring_ctx *ctx = req->ctx;
@@ -4748,7 +4751,7 @@ static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 	req->flags |= sqe_flags & (IOSQE_IO_DRAIN|IOSQE_IO_HARDLINK|
 					IOSQE_ASYNC);
 
-	ret = io_req_set_file(state, req, sqe);
+	ret = io_req_set_file(req, sqe);
 	if (unlikely(ret)) {
 err_req:
 		io_cqring_add_event(req, ret);
@@ -4823,8 +4826,10 @@ static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 /*
  * Batched submission is done, ensure local IO is flushed out.
  */
-static void io_submit_state_end(struct io_submit_state *state)
+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,
@@ -4834,9 +4839,10 @@ static void io_submit_state_end(struct io_submit_state *state)
 /*
  * Start submission side cache.
  */
-static void io_submit_state_start(struct io_submit_state *state,
-				  unsigned int max_ios)
+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;
@@ -4903,7 +4909,6 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 			  struct mm_struct **mm, bool async)
 {
 	struct blk_plug plug;
-	struct io_submit_state state;
 	struct io_kiocb *link = NULL;
 	int i, submitted = 0;
 	bool mm_fault = false;
@@ -4921,7 +4926,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 	if (!percpu_ref_tryget_many(&ctx->refs, nr))
 		return -EAGAIN;
 
-	io_submit_state_start(&state, nr);
+	io_submit_start(ctx, nr);
 	if (nr > IO_PLUG_THRESHOLD)
 		blk_start_plug(&plug);
 
@@ -4932,7 +4937,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 		const struct io_uring_sqe *sqe;
 		struct io_kiocb *req;
 
-		req = io_get_req(ctx, &state);
+		req = io_get_req(ctx);
 		if (unlikely(!req)) {
 			if (!submitted)
 				submitted = -EAGAIN;
@@ -4965,7 +4970,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 		req->needs_fixed_file = async;
 		trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
 						true, async);
-		if (!io_submit_sqe(req, sqe, &state, &link))
+		if (!io_submit_sqe(req, sqe, &link))
 			break;
 	}
 
@@ -4977,7 +4982,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 	if (link)
 		io_queue_link_head(link);
 
-	io_submit_state_end(&state);
+	io_submit_end(ctx);
 	if (nr > IO_PLUG_THRESHOLD)
 		blk_finish_plug(&plug);
 
-- 
2.24.0