[Pvfs2-developers] Re: BMI TCP socket close for sock buf size

[Sam Lang]

On Jul 13, 2006, at 5:22 PM, Pete Wyckoff wrote:

> slang at mcs.anl.gov wrote on Thu, 13 Jul 2006 16:01 -0500:
>> On Jul 13, 2006, at 2:09 PM, Pete Wyckoff wrote:
>>> The
>>> bigger problem that sockbuf sizes should not be global settings, but
>>> rather per-mountpoint, is also not dealt with here.
>>
>> Hmm..but you are only closing the connection to the address from the
>> mountpoint, so unless the same host/port are used for two different
>> mountpoints, we should be ok.
>
> Imagine the case where the client adds two fses (or two mountpoints,
> whatever the right name is).  Connect to config mgr #1, set sockbuf
> sizes #1; connect to config mgr #2, set sockbuf sizes #2.  Later
> connect to an IO server in FS #1 using the sockbuf sizes from #2.
> I'm not particularly motivated to fix this.

Oh I was thinking of the scenario where the second mountpoint would  
cause all the connections from the first mountpoint to be closed  
again.  I guess I could move the TCPBuffer[Send/Receive] options to  
the Filesystem context or inside a NetworkHints
context in the FIlesystem context, but then we would still need to be  
able to keep track of the filesystems in the bmi layer.  Not sure we  
want to do that.

>
>> I would hate to be labeled the style nazi...but are those curly
>> brackets on the same line as the if and case statements? ;-)
>
> They sure are.  You've got 34 other occurences of that to fix in
> bmi.c too.  :)  And another 2350 to fix in the rest of the source.
> I'll point to linux/Documentation/CodingStyle chapter 3 for evidence
> on why I'm right and you're wrong, but agree that I should have
> respected the local custom in that particular file when editing it.

I prefer separate lines for curlies probably from practice more than  
anything, but I'm really just looking for consistency.  This isn't  
one of the hills I'd choose to die on.

-sam

>
> 		-- Pete
>