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

[Steve Wise]

Roland Dreier wrote:
>  > +struct iw_protocol_stats {
>  > +	u64 tcpInSegs;
>  > +	u64 tcpOutSegs;
>  > +	u64 tcpRetransSeg;
>  > +	u64 tcpInErrs;
>  > +	u32 tcpActiveOpens;
>  > +	u32 tcpPassiveOpens;
>  > +	u32 tcpAttemptFails;
>  > +	u32 tcpEstabResets;
>  > +	u32 tcpOutRsts;
>  > +	u32 tcpCurrEstab;
>  > +	u32 tcpRtoMin;
>  > +	u32 tcpRtoMax;
>  > +	u64 ipInReceive;
>  > +	u64 ipInHdrErrors;
>  > +	u64 ipInAddrErrors;
>  > +	u64 ipInUnknownProtos;
>  > +	u64 ipInDiscards;
>  > +	u64 ipInDelivers;
>  > +	u64 ipOutRequests;
>  > +	u64 ipOutDiscards;
>  > +	u64 ipOutNoRoutes;
>  > +	u32 ipReasmTimeout;
>  > +	u32 ipReasmReqds;
>  > +	u32 ipReasmOKs;
>  > +	u32 ipReasmFails;
>  > +};
>
> Two comments about this:
>
>  - how about if we make every field u64 and avoid having to worry about
>    the field width?  I guess cxgb3 made some implementation choices but
>    for example I don't see any good reason that ipInHdrErrors should be
>    64 bits while tcpPassiveOpens is only 32 bits.
>   

This struct was taken from the chelsio MIB, which is  based on the 
standard TCP and IP MIBs.  But the this struct could certainly use 
u64.   And that would simplify the code.  Will do.

>  - similarly let's include all the standard IPv4 and TCP MIB fields, and
>    cxgb3 can just leave the ones it doesn't support as 0.  it looks like
>    this requires zero changes to the cxgb3 implementation anyway.
>   

Ok, but I'm fuzzy on which MIB values we should carry.  For instance, 
some objects in the MIB dump the table of active TCP connections.  I 
don't want to deal with that at this point. 


Stevo.