[OpenAFS] Re: afs.GCPAGs in current releases under Linux (RHEL4/5)

Simon Wilkinson sxw@inf.ed.ac.uk
Sat, 6 Mar 2010 13:17:34 +0000

On 5 Mar 2010, at 19:37, Eric.Hagberg@morganstanley.com wrote:

> On Fri, 5 Mar 2010, Simon Wilkinson wrote:
>> However, if this was happening correctly, Eric should be seeing his
>> system load peak and trough. Performance will be good every 10
>> minutes, and then slowly deteriorate until the next garbage collector
>> run comes along. From what he's reporting, that doesn't seem to be  
>> the
>> case, which is causing me to wonder whether we're actually running
>> GCUserData at all (we should be, as it's independent of the GCPAGS
>> setting)
> If I allow the pag garbage collection (as in the patch I listed in  
> my original email), then I see slow perf degradation, though minor,  
> that cycles every 10 minutes.

Just to update the list. Eric and I continued debugging this off-list,  
and have identified the problem. The keyring destruction function was  
only deleting the first token within each PAG. If you only ever obtain  
tokens for a single cell, this wo't be an issue. However, if you get  
them for more than one cell, all bar one of the token structures are  
leaked, and the performance problems Eric has noted appear as more and  
more redundant objects exist in the kernel.

The fix is available for review at http://gerrit.openafs.org/1524