On Tue, 2024-03-26 at 14:25 +0800, Heng Qi wrote:
> When the dim worker is scheduled, if it fails to acquire the lock,
> dim may not be able to return to the working state later.
>
> For example, the following single queue scenario:
> 1. The dim worker of rxq0 is scheduled, and the dim status is
> changed to DIM_APPLY_NEW_PROFILE;
> 2. The ethtool command is holding rtnl lock;
> 3. Since the rtnl lock is already held, virtnet_rx_dim_work fails
> to acquire the lock and exits;
>
> Then, even if net_dim is invoked again, it cannot work because the
> state is not restored to DIM_START_MEASURE.
>
> Patch has been tested on a VM with 16 NICs, 128 queues per NIC
> (2kq total):
> With dim enabled on all queues, there are many opportunities for
> contention for RTNL lock, and this patch introduces no visible hotspots.
> The dim performance is also stable.
>
> Fixes: 6208799553a8 ("virtio-net: support rx netdim")
> Signed-off-by: Heng Qi <hengqi@xxxxxxxxxxxxxxxxx>
> Acked-by: Jason Wang <jasowang@xxxxxxxxxx>
> ---
> v1->v2:
> - Update commit log. No functional changes.
>
> drivers/net/virtio_net.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index c22d111..0ebe322 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -3563,8 +3563,10 @@ static void virtnet_rx_dim_work(struct work_struct *work)
> struct dim_cq_moder update_moder;
> int i, qnum, err;
>
> - if (!rtnl_trylock())
> + if (!rtnl_trylock()) {
> + schedule_work(&dim->work);
> return;
I'm really scared by this change. VMs are (increasingly) used to run
containers orchestration, which in turns puts a lot of pressure on the
RTNL lock. Any rtnl_trylock+ reschedule may hang for a very long time.
Addressing this kind of issues later becomes _extremely_ painful, see:
https://lore.kernel.org/netdev/20231018154804.420823-1-atenart@xxxxxxxxxx/
I really think a different solution is needed. What about moving
virtnet_send_command() under protection of a new mutex?
I understand it will complicate future hardening works around cvq, but
really rtnl_trylock()/<spin/retry> is bad for the whole system.
Cheers,
Paolo
