[rds-devel] what event is delivered when HCA is hot plugged ?

[Richard Frank]

Posting to rds-dev...

Pradeep, do you also have patches to fix these issues - and can you test 
this solution with hot-plug capability ?

Pradeep wrote:
>
> These are two different issues:
>
> 1. rds_ib_remove_one() is called in case the HCA is going
>   away - due to a PCI fatal error or an admin initiated hotplug
>   remove. rds_ib_remove_one() doesn't seem to cleanup connections
>   that are created on that device. RDS knows the IP addresses that are
>   assigned to the IB device through rds_ibdev->ipaddr_list.
>   So it isn't too difficult to cleanup connections that belong
>   to these IPs.
>
> 2. Currently RDS initiates a reconnect in case of connection
>   failure (an error on the QP). In case of port down, this will
>   happen after the IB transport timeout. I was wondering whether
>   RDS needs to delay it to that time or can it start the reconnect
>   upon a PORT_DOWN event from IB stack. I don't know how OFED stack
>   delivers this event. But I'm sure the HCA driver gets a PORT_DOWN
>   event.
>
> -Pradeep