[dpdk-dev] [PATCH 2/2] net/bonding: fix MAC address when one port resets

[Wei Hu (Xavier)]

Hi, Chas



On 2020/4/4 22:07, Chas Williams wrote:
> The behavior is probably going to be inconsistent. Only one of the calls
> to mac_address_slaves_update() is checked for failure. The two calls
> are in bond_ethdev_lsc_event_callback()
> 
>                  if (internals->active_slave_count < 1) {
>                          mac_address_slaves_update(bonded_eth_dev);
> 
> and in bond_ethdev_start():
> 
>          /* Update all slave devices MACs*/
>          if (mac_address_slaves_update(eth_dev) != 0)
>                  goto out_err;
> 
> So if the bond device is running, you safely fail to the backup device.
> But if you stop and start the bond device, it will then fail to start?
We will make modification as below，and will send patch V2.
After modification，It will then fail to start when failed to setting 
bond device's MAC to the new primary port.

int
mac_address_slaves_update(struct rte_eth_dev *bonded_eth_dev)
{
     <snip>
     bool setted;

     <snip>
     switch (internals->mode) {
     <snip>
     default:
         setted = true;
         for (i = 0; i < internals->slave_count; i++) {
             if (internals->slaves[i].port_id ==
                 internals->current_primary_port) {
                 if (rte_eth_dev_default_mac_addr_set(
                         internals->primary_port,
                         bonded_eth_dev->data->mac_addrs)) {
                             RTE_BOND_LOG(ERR, "Failed to xxx",);
                             setted = false;
                 }
             } else {
                 if (rte_eth_dev_default_mac_addr_set(
                         internals->slaves[i].port_id,
                         &internals->slaves[i].persisted_mac_addr)) {
                             RTE_BOND_LOG(ERR, "Failed to xxx",);
                 }
             }
        }
        if (!setted)
            return -1;
     }
}
I think we can ignore the failure when updating salves's MAC address in 
the mac_address_slaves_update function, because it doesn't affect
the bond's functional characteristics.

> 
> What devices are you bonding together that aren't able to support
> changing their MAC addresses reliably? ixgbe VF devices by any chance?
> One solution is to always assume success or ignore the error from devices
> that you know are not an issue. I agree that we shouldn't early exit
> from mac_address_slaves_update(). I am not sure how to handle the error.
> 
We are running based on huawei's hns3 network engine of kunpeng 920 SoC.
When the driver detects that the hardware status is abnormal, the reset 
process is executed, after hardware reset we will restore the hardware 
configuration including MAC address, and then network engine will be 
able to work normally again. During the hardware reset, the command 
sending to firmware will timeout.

> We probably shouldn't allow you to bond devices that don't support setting
> MAC address. I don't think this check exists currently. That's possibly
> one reason for this check.
> 
> On 2/25/20 4:29 AM, Wei Hu (Xavier) wrote:
>  > From: "Wei Hu (Xavier)" <xavier.huwei at huawei.com>
>  >
>  > Currently, based on a active-backup bond device, in the following 2 
> cases:
>  > 1) The primary port resets. The link status of the primary port changes
>  >     from up to down.
>  > 2) When switching the active port, one slave port resets at the same 
> time.
>  > one slave port changes to the primary port, but the new primary 
> port's MAC
>  > address probably cannot change to the bond device's MAC address. And we
>  > can't continue receive packets whose destination MAC addresses are 
> the same
>  > as the bond devices's MAC address.
>  >
>  > The current bonding PMD driver call mac_address_slaves_update 
> function to
>  > modify the MAC address of all slaves devices. In 
> mac_address_slaves_update
>  > function, the rte_eth_dev_default_mac_addr_set API function is called to
>  > set the MAC address of the slave devices in turn in the for loop 
> statement.
>  >
>  > When one port reset, calling rte_eth_dev_default_mac_addr_set API fails
>  > because the firmware will not respond to the commands from the driver,
>  > and exit the loop, so other slave devices cannot continue to update the
>  > MAC address.
>  >
>  > This patch fixes the issue by avoid exiting the loop when calling
>  > rte_eth_dev_default_mac_addr_set fails.
>  >
>  > Fixes: 2efb58cbab6e ("bond: new link bonding library")
>  > Cc: stable at dpdk.org
>  >
>  > Signed-off-by: Hongbo Zheng <zhenghongbo3 at huawei.com>
>  > Signed-off-by: Wei Hu (Xavier) <xavier.huwei at huawei.com>
>  > Signed-off-by: Chunsong Feng <fengchunsong at huawei.com>
>  > Signed-off-by: Xuan Li <lixuan47 at hisilicon.com>
>  > ---
>  >   drivers/net/bonding/rte_eth_bond_pmd.c | 8 ++++++--
>  >   1 file changed, 6 insertions(+), 2 deletions(-)
>  >
>  > diff --git a/drivers/net/bonding/rte_eth_bond_pmd.c 
> b/drivers/net/bonding/rte_eth_bond_pmd.c
>  > index ddae3518c..ba3f342e7 100644
>  > --- a/drivers/net/bonding/rte_eth_bond_pmd.c
>  > +++ b/drivers/net/bonding/rte_eth_bond_pmd.c
>  > @@ -1502,6 +1502,7 @@ int
>  >   mac_address_slaves_update(struct rte_eth_dev *bonded_eth_dev)
>  >   {
>  >       struct bond_dev_private *internals = 
> bonded_eth_dev->data->dev_private;
>  > +    bool setted;
>  >       int i;
>  >
>  >       /* Update slave devices MAC addresses */
>  > @@ -1529,6 +1530,7 @@ mac_address_slaves_update(struct rte_eth_dev 
> *bonded_eth_dev)
>  >       case BONDING_MODE_TLB:
>  >       case BONDING_MODE_ALB:
>  >       default:
>  > +        setted = true;
>  >           for (i = 0; i < internals->slave_count; i++) {
>  >               if (internals->slaves[i].port_id ==
>  >                       internals->current_primary_port) {
>  > @@ -1537,7 +1539,7 @@ mac_address_slaves_update(struct rte_eth_dev 
> *bonded_eth_dev)
>  >                           bonded_eth_dev->data->mac_addrs)) {
>  >                       RTE_BOND_LOG(ERR, "Failed to update port Id %d 
> MAC address",
>  >                               internals->current_primary_port);
>  > -                    return -1;
>  > +                    setted = false;
>  >                   }
>  >               } else {
>  >                   if (rte_eth_dev_default_mac_addr_set(
>  > @@ -1545,10 +1547,12 @@ mac_address_slaves_update(struct rte_eth_dev 
> *bonded_eth_dev)
>  >                           &internals->slaves[i].persisted_mac_addr)) {
>  >                       RTE_BOND_LOG(ERR, "Failed to update port Id %d 
> MAC address",
>  >                               internals->slaves[i].port_id);
>  > -                    return -1;
>  > +                    setted = false;
>  >                   }
>  >               }
>  >           }
>  > +        if (!setted)
>  > +            return -1;
>  >       }
>  >
>  >       return 0;
>  >