[PATCH] io_uring/rw: end write accounting from ->ki_complete

[Jens Axboe <axboe@kernel.dk>]

Commit b000145e9907 moved both the fsnotify calls and the write
accounting out of the kiocb completion handler and into the
io_req_rw_complete() task_work. However, only the fsnotify part actually
needed to move as it may sleep. Ending the write accounting is just a
percpu_up_read() on the superblock writers sem.

Deferring it is a problem, because it makes dropping SB_FREEZE_WRITE
protection depend on the ring owner getting to running task_work. But
the task may be blocked in freeze_super(), causing it to never get to
that:

  task                             io-wq worker
  --------------------------------------------------------------
  io_write()
    io_kiocb_start_write()         (takes sb_writers, hidden from
                                    lockdep by __sb_writers_release)
    write_iter() -> -EIOCBQUEUED
  ioctl(FS_IOC_SHUTDOWN)
    bdev_freeze()
      freeze_super()
        percpu_down_write()        <- waits for the reader above
                                   io_write()
                                     kiocb_start_write()
                                       percpu_down_read()  <- queued
                                                              behind the
                                                              writer
  <bio completes>
    io_complete_rw()
      queues io_req_rw_complete()  <- never runs, task is in D state

End the write from io_complete_rw() instead, and leave only the fsnotify
calls in task_work.

Reported-by: syzbot+2eb3d983669d3e49d4fa@syzkaller.appspotmail.com
Cc: stable@vger.kernel.org
Fixes: b000145e9907 ("io_uring/rw: defer fsnotify calls to task context")
Signed-off-by: Jens Axboe <axboe@kernel.dk>

---

diff --git a/io_uring/rw.c b/io_uring/rw.c
index 95106dd1d7eb..3e22f294bdf2 100644
--- a/io_uring/rw.c
+++ b/io_uring/rw.c
@@ -517,20 +517,25 @@ static void io_req_end_write(struct io_kiocb *req)
 	}
 }
 
-/*
- * Trigger the notifications after having done some IO, and finish the write
- * accounting, if any.
- */
-static void io_req_io_end(struct io_kiocb *req)
+/* Trigger the notifications after having done some IO. */
+static void io_req_io_notify(struct io_kiocb *req)
 {
 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 
-	if (rw->kiocb.ki_flags & IOCB_WRITE) {
-		io_req_end_write(req);
+	if (rw->kiocb.ki_flags & IOCB_WRITE)
 		fsnotify_modify(req->file);
-	} else {
+	else
 		fsnotify_access(req->file);
-	}
+}
+
+/* Finish write accounting and notify, for inline completions only. */
+static void io_req_io_end(struct io_kiocb *req)
+{
+	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
+
+	if (rw->kiocb.ki_flags & IOCB_WRITE)
+		io_req_end_write(req);
+	io_req_io_notify(req);
 }
 
 static void __io_complete_rw_common(struct io_kiocb *req, long res)
@@ -563,7 +568,7 @@ void io_req_rw_complete(struct io_tw_req tw_req, io_tw_token_t tw)
 {
 	struct io_kiocb *req = tw_req.req;
 
-	io_req_io_end(req);
+	io_req_io_notify(req);
 
 	if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING))
 		req->cqe.flags |= io_put_kbuf(req, max(req->cqe.res, 0), NULL);
@@ -577,6 +582,10 @@ static void io_complete_rw(struct kiocb *kiocb, long res)
 	struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
 	struct io_kiocb *req = cmd_to_io_kiocb(rw);
 
+	/* ring owner may block in freeze_super() before task_work runs */
+	if (kiocb->ki_flags & IOCB_WRITE)
+		io_req_end_write(req);
+
 	__io_complete_rw_common(req, res);
 	io_req_set_res(req, io_fixup_rw_res(req, res), 0);
 	req->io_task_work.func = io_req_rw_complete;

-- 
Jens Axboe