[openib-general] tavor quirks etc (opensm compliance etc)

[Eitan Zahavi]

Hi Or,

Sorry it took me a while.

According to the IBTA spec:
1. In order for MTU and MTUSelector to have any effect their component 
mask bits MUST be set to 1 in the query
2. Behavior of the SM is defined with small "freedom" to choose between 
multiple matching MTU values if they exist.
3. The table below summarizes all options:

Assuming the value M represents the lowest MTU on the path
We denote by M-1 the MTU value one level below M (e.g. 1K if M=2K)
R represents the MTU value in the request. Similarly R-1 is one below R 
and R+1 is one above R.

Query-MTU | Query-Sel | Resp by Spec     | OpenSM Should  | OpenSM Quirk 
w. Tavor End Port
-----------------------------------------------------------------------------------------
UNDEFINED | UNDEFINED | <= M             | M              | min(M,1K)
R         | <         | <= min(R-1, M)   | min(R-1, M)    | min(R-1, M, 1K)
R         | =         | R if M>=R /ERR   | R if M>=R /ERR | R if M>=R /ERR
R         | >         | R < <= M         | R+1 if M>R /ERR| R+1 if M>R /ERR

I have built some test code for making sure OpenSM does what is required.
Apparently it does not. In any case the M is not identical to R it fails 
the request.

I am working on fixing OpenSM.

Any comments are welcome.

EZ

Or Gerlitz wrote:
> Michael S. Tsirkin wrote:
>   
>> I am not yet sure what is best for upstream, so I don't really think we need
>> any RFCs.
>>     
>
>   
>> We'll need data from SM guys on whether MTU selector actually works
>> in SMs, and if not what happens when you enable it.
>>     
>
> Eitan,
>
> Can you please post here the tavor-quirk patch which was integrated into 
> opensm? i can see the ***code*** of the opensm but might make some wrong 
> assumptions or get into wrong understandings as i am not able to see the 
> patch as is.
>
> Or.
>
>
>
>
> _______________________________________________
> openib-general mailing list
> openib-general at openib.org
> http://openib.org/mailman/listinfo/openib-general
>
> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
>