NVMe over fabrics - Caching controller Identity data

NVMe over fabrics - Caching controller Identity data

[Steve Blightman]

As I understand it, the linux initiator caches the controller Identity 
data and some other values.  These only get loaded when it first 
connects, or when the controller is reset.  Now If a target wishes to 
change any of these values, capsule size for instance, it might set the 
new values and restart itself, but when the initiator reconnects to an 
existing controller, it still has the old values. Is there a way for the 
target to force the initiator to reload these values somehow?

NVMe over fabrics - Caching controller Identity data

[Christoph Hellwig]

On Fri, Apr 14, 2017@10:49:10AM -0700, Steve Blightman wrote:
> As I understand it, the linux initiator caches the controller Identity data
> and some other values.  These only get loaded when it first connects, or
> when the controller is reset.  Now If a target wishes to change any of these
> values, capsule size for instance, it might set the new values and restart
> itself, but when the initiator reconnects to an existing controller, it
> still has the old values. Is there a way for the target to force the
> initiator to reload these values somehow?

Identify controller is supposed to be static data for a controller.

If your controller has changed you need to reject the reconnect and
present a new controller.

NVMe over fabrics - Caching controller Identity data

[Sagi Grimberg]

>> As I understand it, the linux initiator caches the controller Identity data
>> and some other values.  These only get loaded when it first connects, or
>> when the controller is reset.  Now If a target wishes to change any of these
>> values, capsule size for instance, it might set the new values and restart
>> itself, but when the initiator reconnects to an existing controller, it
>> still has the old values. Is there a way for the target to force the
>> initiator to reload these values somehow?
>
> Identify controller is supposed to be static data for a controller.
>
> If your controller has changed you need to reject the reconnect and
> present a new controller.

Given that we support dynamic controllers on the host side we need
to issue identify again.

I have some works to centralize this path and there I do re-identify
the controller.