[ofa-general] [PATCH RFCv2 1/2] RDMA: Add protocol statistics attributes to the RDMA sysfs.

[Chien Tung]

> Roland Dreier Wrote:
> > Steve Wise Wrote:
>  > This patch adds a sysfs attribute group called 
> "proto_stats" under  > /sys/class/infiniband/$device/ and 
> populates this group with protocol  > statistics if they 
> exist for a given device.  Currently, only iWARP  > stats are 
> defined, but the code is designed to allow Infiniband 
> protocol  > stats if they become available.  These stats are 
> per-device and more  > importantly -not- per port.
> 
> Making this per-device makes sense I guess given that 
> protocol processing is not bound to a port -- ie in IB we 
> have alternate path migration that can make things switch 
> ports, and for iWARP RNICs presumably an offloaded connection 
> could migrate from one port to another if routing changes 
> (although I doubt we have in-kernel support for this).  And I 
> would guess in most HW implementations, the 
> transport-handling logic has no idea what physical port the 
> packets it's working on is attached to.
> 
> However I see that iw_nes has an ethtool stat for 
> "Retransmission count", which is per-netdev and hence 
> attached to a physical port.  Is this a compromise to work 
> with ethool, or is this really per-port for nes devices?  How 
> does Steve's approach work for you?

The "Retransmission count" is per netdev and we support multiple netdev
per physical port.  We have designed our silicon to support stats at
netdev level and it doesn't make sense for us to aggregate stats from
all netdevs and report at a device level.  Perhaps we can compromise,
Chelsio can give device level stats at /sys/class/infiniband/$device and
for NetEffect, we can provide netdev level stats below it?

Chien