On Tue, 2011-11-01 at 12:33 +0100, Jakub Jankowski wrote:
On Tue, 1 Nov 2011 01:18:38 +0100, Jakub Jankowski wrote:
On Mon, 31 Oct 2011 22:16:21 +0100, Balazs Scheidler wrote:
On Mon, 2011-10-31 at 12:17 +0100, Jakub Jankowski wrote:
2. There are still some leaks; one of which is definitely triggered by HUP. Full valgrind output files are available as: http://toxcorp.com/stuff/syslog-ng-leak/s3.3.1-60b5967-HUP.log http://toxcorp.com/stuff/syslog-ng-leak/s3.3.1-60b5967-noHUP.log [...] Thanks for the detailed diagnosis, there was indeed a leak which can multiply on reload for each open (at the time of the reload), macro-based destination file.
commit 9e7f9258b03c58701be40d40f9708bb3a9c56625
I can confirm that current HEAD does not exhibit this leak any longer:
==25823== LEAK SUMMARY: ==25823== definitely lost: 637 bytes in 14 blocks ==25823== indirectly lost: 0 bytes in 0 blocks ==25823== possibly lost: 13,071 bytes in 107 blocks ==25823== still reachable: 78,991 bytes in 3,277 blocks ==25823== suppressed: 0 bytes in 0 blocks
I've updated my production systems to 9e7f9258b and will observe its behaviour over the next few days (but even 60b5967 looked promising).
I've been running 9e7f9258b for the past ~12 hours and VmRSS grew by about 1MB, which is a huge improvement over what I experienced earlier (roughly 1GB/day).
Thanks for reporting back. Having this nailed I think I go on to prepare 3.3.2 then. -- Bazsi