On Tue, Feb 26, 2019 at 02:01:47PM +0200, Leon Romanovsky wrote:
> @@ -658,7 +658,7 @@ static int nldev_set_doit(struct sk_buff *skb, struct nlmsghdr *nlh,
> return -EINVAL;
>
> index = nla_get_u32(tb[RDMA_NLDEV_ATTR_DEV_INDEX]);
> - device = ib_device_get_by_index(index);
> + device = ib_device_get_by_index(sock_net(skb->sk), index);
> if (!device)
> return -EINVAL;
>
> @@ -685,6 +685,14 @@ static int _nldev_get_dumpit(struct ib_device *device,
> if (idx < start)
> return 0;
>
> + if (!ib_device_try_get(device))
> + return 0;
Why is this added?
_nldev_get_dumpit is only called from ib_enum_all_devs which is
holding the devices_rwsem
I think the test for rdma_dev_access_netns should just be part
of ib_enum_all_devs(), not in each of the callbacks.
> + if (!rdma_dev_access_netns(device, sock_net(skb->sk))) {
> + ib_device_put(device);
> + return 0;
This placement is wrong, it should be in ib_enum_all_devs() and 'idx'
should not include devices outside the namespace. ie should should not
be holes in the idx numbering for devices which are invsible.
Jason
