As diagnosed by Song Liu, ndo_poll_controller() can be very dangerous on loaded hosts, since the cpu calling ndo_poll_controller() might steal all NAPI contexts (for all RX/TX queues of the NIC).
This capture, showing one ksoftirqd eating all cycles can last for unlimited amount of time, since one cpu is generally not able to drain all the queues under load. It seems that all networking drivers that do use NAPI for their TX completions, should not provide a ndo_poll_controller() : Most NAPI drivers have netpoll support already handled in core networking stack, since netpoll_poll_dev() uses poll_napi(dev) to iterate through registered NAPI contexts for a device. First patch is a fix in poll_one_napi(). Then following patches take care of ten drivers. Eric Dumazet (11): netpoll: do not test NAPI_STATE_SCHED in poll_one_napi() hinic: remove ndo_poll_controller ehea: remove ndo_poll_controller net: hns: remove ndo_poll_controller virtio_net: remove ndo_poll_controller qlcnic: remove ndo_poll_controller qlogic: netxen: remove ndo_poll_controller net: ena: remove ndo_poll_controller sfc: remove ndo_poll_controller sfc-falcon: remove ndo_poll_controller ibmvnic: remove ndo_poll_controller drivers/net/ethernet/amazon/ena/ena_netdev.c | 22 --------- drivers/net/ethernet/hisilicon/hns/hns_enet.c | 18 -------- .../net/ethernet/huawei/hinic/hinic_main.c | 20 --------- drivers/net/ethernet/ibm/ehea/ehea_main.c | 14 ------ drivers/net/ethernet/ibm/ibmvnic.c | 16 ------- .../ethernet/qlogic/netxen/netxen_nic_main.c | 23 ---------- .../net/ethernet/qlogic/qlcnic/qlcnic_main.c | 45 ------------------- drivers/net/ethernet/sfc/efx.c | 26 ----------- drivers/net/ethernet/sfc/falcon/efx.c | 26 ----------- drivers/net/virtio_net.c | 14 ------ net/core/netpoll.c | 20 +-------- 11 files changed, 1 insertion(+), 243 deletions(-) -- 2.19.0.605.g01d371f741-goog