Re: [PATCH net-next v08 7/9] hinic3: Add adaptive IRQ coalescing with DIM

[Jakub Kicinski]

AI code review points out:

> @@ -150,6 +236,9 @@ int hinic3_qps_irq_init(struct net_device *netdev)
>  			goto err_release_irqs;
>  		}
>
> +		INIT_WORK(&irq_cfg->rxq->dim.work, hinic3_rx_dim_work);
> +		irq_cfg->rxq->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_CQE;
> +
>  		hinic3_set_msix_auto_mask_state(nic_dev->hwdev,
>  						irq_cfg->msix_entry_idx,
>  						HINIC3_SET_MSIX_AUTO_MASK);
> @@ -164,6 +253,9 @@ int hinic3_qps_irq_init(struct net_device *netdev)
>  		q_id--;
>  		irq_cfg = &nic_dev->q_params.irq_cfg[q_id];
>  		qp_del_napi(irq_cfg);
> +
> +		disable_work_sync(&irq_cfg->rxq->dim.work);
> +
>  		hinic3_set_msix_state(nic_dev->hwdev, irq_cfg->msix_entry_idx,
>  				      HINIC3_MSIX_DISABLE);

The error path in hinic3_qps_irq_init() calls disable_work_sync() to cancel
the DIM work before cleanup. However, hinic3_qps_irq_uninit() does not have
a corresponding cancel_work_sync() or disable_work_sync() call.

The DIM work is scheduled via net_dim()->schedule_work() from hinic3_poll().
If the interface is closed while DIM work is pending or running,
hinic3_rx_dim_work() could access rxq->netdev and rxq->q_id after the
queues have been torn down in hinic3_close_channel()->hinic3_qps_irq_uninit().

Should hinic3_qps_irq_uninit() call cancel_work_sync() for the DIM work
to prevent a potential use-after-free?