Under load, the RX side of the mscan driver can get stuck while TX still works. Restarting the interface locks up the system. This behaviour could be reproduced reliably on a MPC5121e based system.
The patch fixes the return value of the NAPI polling function (should be the number of processed packets, not constant 1) and the condition under which IRQs are enabled again after polling is finished.
With this patch, no more lockups were observed over a test period of ten days.
Signed-off-by: Florian Faber faber@faberman.de --- diff -uprN a/drivers/net/can/mscan/mscan.c b/drivers/net/can/mscan/mscan.c --- a/drivers/net/can/mscan/mscan.c +++ b/drivers/net/can/mscan/mscan.c @@ -381,10 +381,9 @@ static int mscan_rx_poll(struct napi_str struct net_device *dev = napi->dev; struct mscan_regs __iomem *regs = priv->reg_base; struct net_device_stats *stats = &dev->stats; - int npackets = 0; - int ret = 1; + int work_done = 0; struct sk_buff *skb; struct can_frame *frame; u8 canrflg;
- while (npackets < quota) { + while (work_done < quota) { canrflg = in_8(®s->canrflg); if (!(canrflg & (MSCAN_RXF | MSCAN_ERR_IF))) @@ -408,15 +409,16 @@ static int mscan_rx_poll(struct napi_str
stats->rx_packets++; stats->rx_bytes += frame->can_dlc; - npackets++; + work_done++; netif_receive_skb(skb); }
- if (!(in_8(®s->canrflg) & (MSCAN_RXF | MSCAN_ERR_IF))) { - napi_complete(&priv->napi); - clear_bit(F_RX_PROGRESS, &priv->flags); - if (priv->can.state < CAN_STATE_BUS_OFF) - out_8(®s->canrier, priv->shadow_canrier); - ret = 0; + if (work_done < quota) { + if (likely(napi_complete_done(&priv->napi, work_done))) { + clear_bit(F_RX_PROGRESS, &priv->flags); + if (priv->can.state < CAN_STATE_BUS_OFF) + out_8(®s->canrier, priv->shadow_canrier); + } } - return ret; + + return work_done; }