io-wq: split hashing and enqueueing

It's a preparation patch removing io_wq_enqueue_hashed(), which
now should be done by io_wq_hash_work() + io_wq_enqueue().

Also, set hash value for dependant works, and do it as late as possible,
because req->file can be unavailable before. This hash will be ignored
by io-wq.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

fs/io-wq.c

@@ -385,7 +385,7 @@ static struct io_wq_work *io_get_next_work(struct io_wqe *wqe, unsigned *hash)
 		work = container_of(node, struct io_wq_work, list);
 
 		/* not hashed, can run anytime */
-		if (!(work->flags & IO_WQ_WORK_HASHED)) {
+		if (!io_wq_is_hashed(work)) {
 			wq_node_del(&wqe->work_list, node, prev);
 			return work;
 		}
@@ -795,19 +795,15 @@ void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 }
 
 /*
- * Enqueue work, hashed by some key. Work items that hash to the same value
+ * Work items that hash to the same value will not be done in parallel.
- * will not be done in parallel. Used to limit concurrent writes, generally
+ * Used to limit concurrent writes, generally hashed by inode.
- * hashed by inode.
  */
-void io_wq_enqueue_hashed(struct io_wq *wq, struct io_wq_work *work, void *val)
+void io_wq_hash_work(struct io_wq_work *work, void *val)
 {
-	struct io_wqe *wqe = wq->wqes[numa_node_id()];
+	unsigned int bit;
-	unsigned bit;
-
 
 	bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 	work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
-	io_wqe_enqueue(wqe, work);
 }
 
 static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data)

fs/io-wq.h

@@ -94,7 +94,12 @@ bool io_wq_get(struct io_wq *wq, struct io_wq_data *data);
 void io_wq_destroy(struct io_wq *wq);
 
 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work);
-void io_wq_enqueue_hashed(struct io_wq *wq, struct io_wq_work *work, void *val);
+void io_wq_hash_work(struct io_wq_work *work, void *val);
+
+static inline bool io_wq_is_hashed(struct io_wq_work *work)
+{
+	return work->flags & IO_WQ_WORK_HASHED;
+}
 
 void io_wq_cancel_all(struct io_wq *wq);
 enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork);

fs/io_uring.c

@@ -1040,15 +1040,14 @@ static inline void io_req_work_drop_env(struct io_kiocb *req)
 	}
 }
 
-static inline bool io_prep_async_work(struct io_kiocb *req,
+static inline void io_prep_async_work(struct io_kiocb *req,
 				      struct io_kiocb **link)
 {
 	const struct io_op_def *def = &io_op_defs[req->opcode];
-	bool do_hashed = false;
 
 	if (req->flags & REQ_F_ISREG) {
 		if (def->hash_reg_file)
-			do_hashed = true;
+			io_wq_hash_work(&req->work, file_inode(req->file));
 	} else {
 		if (def->unbound_nonreg_file)
 			req->work.flags |= IO_WQ_WORK_UNBOUND;
@@ -1057,25 +1056,18 @@ static inline bool io_prep_async_work(struct io_kiocb *req,
 	io_req_work_grab_env(req, def);
 
 	*link = io_prep_linked_timeout(req);
-	return do_hashed;
 }
 
 static inline void io_queue_async_work(struct io_kiocb *req)
 {
 	struct io_ring_ctx *ctx = req->ctx;
 	struct io_kiocb *link;
-	bool do_hashed;
 
-	do_hashed = io_prep_async_work(req, &link);
+	io_prep_async_work(req, &link);
 
-	trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
+	trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
-					req->flags);
+					&req->work, req->flags);
-	if (!do_hashed) {
+	io_wq_enqueue(ctx->io_wq, &req->work);
-		io_wq_enqueue(ctx->io_wq, &req->work);
-	} else {
-		io_wq_enqueue_hashed(ctx->io_wq, &req->work,
-					file_inode(req->file));
-	}
 
 	if (link)
 		io_queue_linked_timeout(link);
@@ -1582,6 +1574,10 @@ static void io_link_work_cb(struct io_wq_work **workptr)
 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
 {
 	struct io_kiocb *link;
+	const struct io_op_def *def = &io_op_defs[nxt->opcode];
+
+	if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
+		io_wq_hash_work(&nxt->work, file_inode(nxt->file));
 
 	*workptr = &nxt->work;
 	link = io_prep_linked_timeout(nxt);