[ofa-general] Re: [PATCH] opensm/osm_switch.h: use updated LFT for routing

[Yevgeny Kliteynik]

Sasha,

Sasha Khapyorsky wrote:
> Hi Yevgeny,
> 
> On 13:58 Thu 20 Nov     , Yevgeny Kliteynik wrote:
>> Function osm_switch_get_port_by_lid() was using the switch's
>> LFT, so this LFT might not be updated to recent routing.
> 
> I guess it could be only with 'subnet_initialization_error' flag up
> (failed LinFwdTbl set will trigger this flag).
>> I think that this was also relevant before the LFT simplification.
> 
> Yes, logically it should be so, but...
> 
>> One immediate outcome of this bug is opensm.fdbs file - when it
>> is dumped from the switch LFT (and not from lft_buf),
> 
> Why this bug is triggered only now?

I had sometimes errors in simulations, and after aome analysis
I decided that they are timing problems with the tests.
Now that I did some stress testing of ucast cache, I started
to see more of these errors.

>> it sometimes
>> doesn't match the lst file.
> 
> What this "sometimes" mean? I think the case should be investigated
> deeper. By such patch we are just trying to hide a possible issue.
> 
> As far as I understand opensm.fdbs (and other routing dump) are
> generated only after all LinFwdTbl responses are arrived, when some of
> them failed 'subnet_initialization_error' flag is up and OpenSM will
> resweep. If so why is 'opensm.fdbs' broken? It is not immediately
> clear for me.

I didn't see 'subnet_initialization_error' in such cases.
Anyway, here's what I can do: at the end of each ucast_mgr_process
I'll compare lft and lft_buf (something that the other patch is
doing, the one that frees lft_buf), and if there is a difference,
then we have a problem. In not - then I'll look for the cause
elsewhere.

-- Yevgeny

> Sasha
> 
>> Signed-off-by: Yevgeny Kliteynik <kliteyn at dev.mellanox.co.il>
>> ---
>>  opensm/include/opensm/osm_switch.h |    6 +++++-
>>  1 files changed, 5 insertions(+), 1 deletions(-)
>>
>> diff --git a/opensm/include/opensm/osm_switch.h b/opensm/include/opensm/osm_switch.h
>> index caa0bc5..f06931c 100644
>> --- a/opensm/include/opensm/osm_switch.h
>> +++ b/opensm/include/opensm/osm_switch.h
>> @@ -411,7 +411,11 @@ osm_switch_get_port_by_lid(IN const osm_switch_t * const p_sw,
>>  {
>>  	if (lid_ho == 0 || lid_ho > IB_LID_UCAST_END_HO)
>>  		return OSM_NO_PATH;
>> -	return p_sw->lft[lid_ho];
>> +
>> +	if (p_sw->lft_buf)
>> +		return p_sw->lft_buf[lid_ho];
>> +	else
>> +		return p_sw->lft[lid_ho];
>>  }
>>  /*
>>  * PARAMETERS
>> -- 
>> 1.5.1.4
>>
>>
>