io_uring: enable req cache for IRQ driven IO

This is the last class of requests that cannot utilize the req alloc
cache. Add a per-ctx req cache that is protected by the completion_lock,
and refill our submit side cache when it gets over our batch count.
Signed-off-by: Jens Axboe <axboe@kernel.dk>

fs/io_uring.c

@@ -272,7 +272,11 @@ struct io_sq_data {
 struct io_comp_state {
 	struct io_kiocb		*reqs[IO_COMPL_BATCH];
 	unsigned int		nr;
+	unsigned int		locked_free_nr;
+	/* inline/task_work completion list, under ->uring_lock */
 	struct list_head	free_list;
+	/* IRQ completion list, under ->completion_lock */
+	struct list_head	locked_free_list;
 };
 
 struct io_submit_state {
@@ -1033,6 +1037,9 @@ static void io_cqring_fill_event(struct io_kiocb *req, long res);
 static void io_put_req(struct io_kiocb *req);
 static void io_put_req_deferred(struct io_kiocb *req, int nr);
 static void io_double_put_req(struct io_kiocb *req);
+static void io_dismantle_req(struct io_kiocb *req);
+static void io_put_task(struct task_struct *task, int nr);
+static void io_queue_next(struct io_kiocb *req);
 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
 static void __io_queue_linked_timeout(struct io_kiocb *req);
 static void io_queue_linked_timeout(struct io_kiocb *req);
@@ -1353,6 +1360,7 @@ static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 	INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
 	init_llist_head(&ctx->rsrc_put_llist);
 	INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
+	INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
 	return ctx;
 err:
 	kfree(ctx->cancel_hash);
@@ -1908,8 +1916,8 @@ static void io_cqring_fill_event(struct io_kiocb *req, long res)
 	__io_cqring_fill_event(req, res, 0);
 }
 
-static void io_req_complete_post(struct io_kiocb *req, long res,
-				 unsigned int cflags)
+static inline void io_req_complete_post(struct io_kiocb *req, long res,
+					unsigned int cflags)
 {
 	struct io_ring_ctx *ctx = req->ctx;
 	unsigned long flags;
@@ -1917,16 +1925,26 @@ static void io_req_complete_post(struct io_kiocb *req, long res,
 	spin_lock_irqsave(&ctx->completion_lock, flags);
 	__io_cqring_fill_event(req, res, cflags);
 	io_commit_cqring(ctx);
+	/*
+	 * If we're the last reference to this request, add to our locked
+	 * free_list cache.
+	 */
+	if (refcount_dec_and_test(&req->refs)) {
+		struct io_comp_state *cs = &ctx->submit_state.comp;
+
+		io_dismantle_req(req);
+		io_put_task(req->task, 1);
+		list_add(&req->compl.list, &cs->locked_free_list);
+		cs->locked_free_nr++;
+	} else
+		req = NULL;
 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 
 	io_cqring_ev_posted(ctx);
-}
-
-static inline void io_req_complete_nostate(struct io_kiocb *req, long res,
-					   unsigned int cflags)
-{
-	io_req_complete_post(req, res, cflags);
-	io_put_req(req);
+	if (req) {
+		io_queue_next(req);
+		percpu_ref_put(&ctx->refs);
+	}
 }
 
 static void io_req_complete_state(struct io_kiocb *req, long res,
@@ -1944,7 +1962,7 @@ static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
 	if (issue_flags & IO_URING_F_COMPLETE_DEFER)
 		io_req_complete_state(req, res, cflags);
 	else
-		io_req_complete_nostate(req, res, cflags);
+		io_req_complete_post(req, res, cflags);
 }
 
 static inline void io_req_complete(struct io_kiocb *req, long res)
@@ -1952,12 +1970,26 @@ static inline void io_req_complete(struct io_kiocb *req, long res)
 	__io_req_complete(req, 0, res, 0);
 }
 
-static bool io_flush_cached_reqs(struct io_submit_state *state)
+static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
 {
+	struct io_submit_state *state = &ctx->submit_state;
+	struct io_comp_state *cs = &state->comp;
 	struct io_kiocb *req = NULL;
 
-	while (!list_empty(&state->comp.free_list)) {
-		req = list_first_entry(&state->comp.free_list, struct io_kiocb,
+	/*
+	 * If we have more than a batch's worth of requests in our IRQ side
+	 * locked cache, grab the lock and move them over to our submission
+	 * side cache.
+	 */
+	if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
+		spin_lock_irq(&ctx->completion_lock);
+		list_splice_init(&cs->locked_free_list, &cs->free_list);
+		cs->locked_free_nr = 0;
+		spin_unlock_irq(&ctx->completion_lock);
+	}
+
+	while (!list_empty(&cs->free_list)) {
+		req = list_first_entry(&cs->free_list, struct io_kiocb,
 					compl.list);
 		list_del(&req->compl.list);
 		state->reqs[state->free_reqs++] = req;
@@ -1978,7 +2010,7 @@ static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
 		gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
 		int ret;
 
-		if (io_flush_cached_reqs(state))
+		if (io_flush_cached_reqs(ctx))
 			goto got_req;
 
 		ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
@@ -8748,14 +8780,12 @@ static void io_destroy_buffers(struct io_ring_ctx *ctx)
 	idr_destroy(&ctx->io_buffer_idr);
 }
 
-static void io_req_cache_free(struct io_ring_ctx *ctx)
+static void io_req_cache_free(struct list_head *list)
 {
-	struct io_comp_state *cs = &ctx->submit_state.comp;
-
-	while (!list_empty(&cs->free_list)) {
+	while (!list_empty(list)) {
 		struct io_kiocb *req;
 
-		req = list_first_entry(&cs->free_list, struct io_kiocb, compl.list);
+		req = list_first_entry(list, struct io_kiocb, compl.list);
 		list_del(&req->compl.list);
 		kmem_cache_free(req_cachep, req);
 	}
@@ -8803,7 +8833,8 @@ static void io_ring_ctx_free(struct io_ring_ctx *ctx)
 	free_uid(ctx->user);
 	put_cred(ctx->creds);
 	kfree(ctx->cancel_hash);
-	io_req_cache_free(ctx);
+	io_req_cache_free(&ctx->submit_state.comp.free_list);
+	io_req_cache_free(&ctx->submit_state.comp.locked_free_list);
 	kfree(ctx);
 }