io_uring: fix REQ_F_COMP_LOCKED by killing it

REQ_F_COMP_LOCKED is used and implemented in a buggy way. The problem is
that the flag is set before io_put_req() but not cleared after, and if
that wasn't the final reference, the request will be freed with the flag
set from some other context, which may not hold a spinlock. That means
possible races with removing linked timeouts and unsynchronised
completion (e.g. access to CQ).

Instead of fixing REQ_F_COMP_LOCKED, kill the flag and use
task_work_add() to move such requests to a fresh context to free from
it, as was done with __io_free_req_finish().

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

1 files changed, 69 insertions, 80 deletions

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 0680fa385353..641d869d96ee 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -574,7 +574,6 @@ enum {
 	REQ_F_NOWAIT_BIT,
 	REQ_F_LINK_TIMEOUT_BIT,
 	REQ_F_ISREG_BIT,
-	REQ_F_COMP_LOCKED_BIT,
 	REQ_F_NEED_CLEANUP_BIT,
 	REQ_F_POLLED_BIT,
 	REQ_F_BUFFER_SELECTED_BIT,
@@ -613,8 +612,6 @@ enum {
 	REQ_F_LINK_TIMEOUT	= BIT(REQ_F_LINK_TIMEOUT_BIT),
 	/* regular file */
 	REQ_F_ISREG		= BIT(REQ_F_ISREG_BIT),
-	/* completion under lock */
-	REQ_F_COMP_LOCKED	= BIT(REQ_F_COMP_LOCKED_BIT),
 	/* needs cleanup */
 	REQ_F_NEED_CLEANUP	= BIT(REQ_F_NEED_CLEANUP_BIT),
 	/* already went through poll handler */
@@ -963,8 +960,8 @@ static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
 			     struct io_comp_state *cs);
 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_double_put_req(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);
@@ -1316,9 +1313,8 @@ static void io_kill_timeout(struct io_kiocb *req)
 		atomic_set(&req->ctx->cq_timeouts,
 			atomic_read(&req->ctx->cq_timeouts) + 1);
 		list_del_init(&req->timeout.list);
-		req->flags |= REQ_F_COMP_LOCKED;
 		io_cqring_fill_event(req, 0);
-		io_put_req(req);
+		io_put_req_deferred(req, 1);
 	}
 }
 
@@ -1369,8 +1365,7 @@ static void __io_queue_deferred(struct io_ring_ctx *ctx)
 		if (link) {
 			__io_queue_linked_timeout(link);
 			/* drop submission reference */
-			link->flags |= REQ_F_COMP_LOCKED;
-			io_put_req(link);
+			io_put_req_deferred(link, 1);
 		}
 		kfree(de);
 	} while (!list_empty(&ctx->defer_list));
@@ -1597,13 +1592,19 @@ static void io_submit_flush_completions(struct io_comp_state *cs)
 		req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
 		list_del(&req->compl.list);
 		__io_cqring_fill_event(req, req->result, req->compl.cflags);
-		if (!(req->flags & REQ_F_LINK_HEAD)) {
-			req->flags |= REQ_F_COMP_LOCKED;
-			io_put_req(req);
-		} else {
+
+		/*
+		 * io_free_req() doesn't care about completion_lock unless one
+		 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
+		 * because of a potential deadlock with req->work.fs->lock
+		 */
+		if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
+				 |REQ_F_WORK_INITIALIZED)) {
 			spin_unlock_irq(&ctx->completion_lock);
 			io_put_req(req);
 			spin_lock_irq(&ctx->completion_lock);
+		} else {
+			io_put_req(req);
 		}
 	}
 	io_commit_cqring(ctx);
@@ -1702,10 +1703,14 @@ static void io_dismantle_req(struct io_kiocb *req)
 	io_req_clean_work(req);
 }
 
-static void __io_free_req_finish(struct io_kiocb *req)
+static void __io_free_req(struct io_kiocb *req)
 {
-	struct io_uring_task *tctx = req->task->io_uring;
-	struct io_ring_ctx *ctx = req->ctx;
+	struct io_uring_task *tctx;
+	struct io_ring_ctx *ctx;
+
+	io_dismantle_req(req);
+	tctx = req->task->io_uring;
+	ctx = req->ctx;
 
 	atomic_long_inc(&tctx->req_complete);
 	if (tctx->in_idle)
@@ -1719,33 +1724,6 @@ static void __io_free_req_finish(struct io_kiocb *req)
 	percpu_ref_put(&ctx->refs);
 }
 
-static void io_req_task_file_table_put(struct callback_head *cb)
-{
-	struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
-
-	io_dismantle_req(req);
-	__io_free_req_finish(req);
-}
-
-static void __io_free_req(struct io_kiocb *req)
-{
-	if (!(req->flags & REQ_F_COMP_LOCKED)) {
-		io_dismantle_req(req);
-		__io_free_req_finish(req);
-	} else {
-		int ret;
-
-		init_task_work(&req->task_work, io_req_task_file_table_put);
-		ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
-		if (unlikely(ret)) {
-			struct task_struct *tsk;
-
-			tsk = io_wq_get_task(req->ctx->io_wq);
-			task_work_add(tsk, &req->task_work, 0);
-		}
-	}
-}
-
 static bool io_link_cancel_timeout(struct io_kiocb *req)
 {
 	struct io_timeout_data *io = req->async_data;
@@ -1754,11 +1732,10 @@ static bool io_link_cancel_timeout(struct io_kiocb *req)
 
 	ret = hrtimer_try_to_cancel(&io->timer);
 	if (ret != -1) {
-		req->flags |= REQ_F_COMP_LOCKED;
 		io_cqring_fill_event(req, -ECANCELED);
 		io_commit_cqring(ctx);
 		req->flags &= ~REQ_F_LINK_HEAD;
-		io_put_req(req);
+		io_put_req_deferred(req, 1);
 		return true;
 	}
 
@@ -1785,17 +1762,12 @@ static bool __io_kill_linked_timeout(struct io_kiocb *req)
 static void io_kill_linked_timeout(struct io_kiocb *req)
 {
 	struct io_ring_ctx *ctx = req->ctx;
+	unsigned long flags;
 	bool wake_ev;
 
-	if (!(req->flags & REQ_F_COMP_LOCKED)) {
-		unsigned long flags;
-
-		spin_lock_irqsave(&ctx->completion_lock, flags);
-		wake_ev = __io_kill_linked_timeout(req);
-		spin_unlock_irqrestore(&ctx->completion_lock, flags);
-	} else {
-		wake_ev = __io_kill_linked_timeout(req);
-	}
+	spin_lock_irqsave(&ctx->completion_lock, flags);
+	wake_ev = __io_kill_linked_timeout(req);
+	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 
 	if (wake_ev)
 		io_cqring_ev_posted(ctx);
@@ -1835,27 +1807,29 @@ static void __io_fail_links(struct io_kiocb *req)
 		trace_io_uring_fail_link(req, link);
 
 		io_cqring_fill_event(link, -ECANCELED);
-		link->flags |= REQ_F_COMP_LOCKED;
-		__io_double_put_req(link);
+
+		/*
+		 * It's ok to free under spinlock as they're not linked anymore,
+		 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
+		 * work.fs->lock.
+		 */
+		if (link->flags & REQ_F_WORK_INITIALIZED)
+			io_put_req_deferred(link, 2);
+		else
+			io_double_put_req(link);
 	}
 
 	io_commit_cqring(ctx);
-	io_cqring_ev_posted(ctx);
 }
 
 static void io_fail_links(struct io_kiocb *req)
 {
 	struct io_ring_ctx *ctx = req->ctx;
+	unsigned long flags;
 
-	if (!(req->flags & REQ_F_COMP_LOCKED)) {
-		unsigned long flags;
-
-		spin_lock_irqsave(&ctx->completion_lock, flags);
-		__io_fail_links(req);
-		spin_unlock_irqrestore(&ctx->completion_lock, flags);
-	} else {
-		__io_fail_links(req);
-	}
+	spin_lock_irqsave(&ctx->completion_lock, flags);
+	__io_fail_links(req);
+	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 
 	io_cqring_ev_posted(ctx);
 }
@@ -2069,6 +2043,34 @@ static void io_put_req(struct io_kiocb *req)
 		io_free_req(req);
 }
 
+static void io_put_req_deferred_cb(struct callback_head *cb)
+{
+	struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
+
+	io_free_req(req);
+}
+
+static void io_free_req_deferred(struct io_kiocb *req)
+{
+	int ret;
+
+	init_task_work(&req->task_work, io_put_req_deferred_cb);
+	ret = io_req_task_work_add(req, true);
+	if (unlikely(ret)) {
+		struct task_struct *tsk;
+
+		tsk = io_wq_get_task(req->ctx->io_wq);
+		task_work_add(tsk, &req->task_work, 0);
+		wake_up_process(tsk);
+	}
+}
+
+static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
+{
+	if (refcount_sub_and_test(refs, &req->refs))
+		io_free_req_deferred(req);
+}
+
 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
 {
 	struct io_kiocb *nxt;
@@ -2085,17 +2087,6 @@ static struct io_wq_work *io_steal_work(struct io_kiocb *req)
 	return nxt ? &nxt->work : NULL;
 }
 
-/*
- * Must only be used if we don't need to care about links, usually from
- * within the completion handling itself.
- */
-static void __io_double_put_req(struct io_kiocb *req)
-{
-	/* drop both submit and complete references */
-	if (refcount_sub_and_test(2, &req->refs))
-		__io_free_req(req);
-}
-
 static void io_double_put_req(struct io_kiocb *req)
 {
 	/* drop both submit and complete references */
@@ -5127,9 +5118,8 @@ static bool io_poll_remove_one(struct io_kiocb *req)
 	if (do_complete) {
 		io_cqring_fill_event(req, -ECANCELED);
 		io_commit_cqring(req->ctx);
-		req->flags |= REQ_F_COMP_LOCKED;
 		req_set_fail_links(req);
-		io_put_req(req);
+		io_put_req_deferred(req, 1);
 	}
 
 	return do_complete;
@@ -5311,9 +5301,8 @@ static int __io_timeout_cancel(struct io_kiocb *req)
 	list_del_init(&req->timeout.list);
 
 	req_set_fail_links(req);
-	req->flags |= REQ_F_COMP_LOCKED;
 	io_cqring_fill_event(req, -ECANCELED);
-	io_put_req(req);
+	io_put_req_deferred(req, 1);
 	return 0;
 }