nbd: add a flush_workqueue in nbd_start_device
authorSun Ke <sunke32@huawei.com>
Wed, 22 Jan 2020 03:18:57 +0000 (11:18 +0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 24 Feb 2020 07:36:31 +0000 (08:36 +0100)
[ Upstream commit 5c0dd228b5fc30a3b732c7ae2657e0161ec7ed80 ]

When kzalloc fail, may cause trying to destroy the
workqueue from inside the workqueue.

If num_connections is m (2 < m), and NO.1 ~ NO.n
(1 < n < m) kzalloc are successful. The NO.(n + 1)
failed. Then, nbd_start_device will return ENOMEM
to nbd_start_device_ioctl, and nbd_start_device_ioctl
will return immediately without running flush_workqueue.
However, we still have n recv threads. If nbd_release
run first, recv threads may have to drop the last
config_refs and try to destroy the workqueue from
inside the workqueue.

To fix it, add a flush_workqueue in nbd_start_device.

Fixes: e9e006f5fcf2 ("nbd: fix max number of supported devs")
Signed-off-by: Sun Ke <sunke32@huawei.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Sasha Levin <sashal@kernel.org>
drivers/block/nbd.c

index b4607dd..7818190 100644 (file)
@@ -1265,6 +1265,16 @@ static int nbd_start_device(struct nbd_device *nbd)
                args = kzalloc(sizeof(*args), GFP_KERNEL);
                if (!args) {
                        sock_shutdown(nbd);
+                       /*
+                        * If num_connections is m (2 < m),
+                        * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
+                        * But NO.(n + 1) failed. We still have n recv threads.
+                        * So, add flush_workqueue here to prevent recv threads
+                        * dropping the last config_refs and trying to destroy
+                        * the workqueue from inside the workqueue.
+                        */
+                       if (i)
+                               flush_workqueue(nbd->recv_workq);
                        return -ENOMEM;
                }
                sk_set_memalloc(config->socks[i]->sock->sk);