On 10/26/2011 06:59 PM, Amos Jeffries wrote: > On Thu, 27 Oct 2011 01:59:41 +0400, Dmitry Kurochkin wrote: >> On Thu, 27 Oct 2011 10:34:34 +1300, Amos Jeffries >> <[email protected]> wrote: >>> On Thu, 27 Oct 2011 01:25:59 +0400, Dmitry Kurochkin wrote: >>> > Hi all. >>> > >>> > There has been no discussion on this thread for a while. The latest >>> > patch version (attached) fixes all the issues from Amos and Alex I >>> > know >>> > of. So I would like to get an approval for the patch or hints at >>> > what >>> > else needs to be fixed. >>> > >>> > The patch does not change the way unlinkd is started. But after the >>> > changes, unlinkd is started iff it is needed. Whether unlinkd is >>> > needed >>> > is determined by both cache dir and DiskIO strategy. Unlinkd may be >>> > started during reconfiguration if a cache dir that needs it was added >>> > or >>> > DiskIO strategy changed. >>> > >>> > There was discussion about starting unlinkd on demand. But the >>> > consensus seems to be that it Squid needs substantial changes to >>> > minimize fork() performance penalty. So while we want to implement >>> > this >>> > long term, for now we should start unlinkd on startup and >>> > reconfiguration. >>> > >>> > Please let me know if I missed any comments or issues with the patch. >>> > >>> > Regards, >>> > Dmitry >>> >>> >>> This seems fine to go in for the bug fix. The rest was us going off on >>> a tangent. >>> >>> The only worry I have left is the dependency on IamWorkerProcess() in >>> src/fs/ufs/store_dir_ufs.cc. >>> >>> If you have checked that SMP with Diskers is not broken by that check >>> then this is fine to go in now. >>> >> >> Yes. IamWorkerProcess() prevents UFS from starting unlinkd in >> coordinator, master and disker processes (though in non-daemon mode, >> master process is worker and). This is what we want. >> > > Then I'm happy for it to go in now. > > +1.
Committed to trunk as r11826. Thank you, Alex.
