Re: ioflush kernel thread chewing CPU time

[Andrew Doran <ad%NetBSD.org@localhost>]

On Tue, Mar 31, 2009 at 07:20:18PM +1100, Simon Burge wrote:

> Andrew Doran wrote:
> 
> > On Wed, Mar 25, 2009 at 12:56:59PM +1100, Simon Burge wrote:
> > 
> > > Simon Burge wrote:
> > > 
> > > > [ profiling working with enami's patch ]
> > > 
> > > I've got two 30-second profiles while I'm seeing ioflush using a lot
> > > of CPU time.  With a profiled kernel I'm seeing ioflush use about 30-35%
> > > instead of the "normal" 20%.  The first of these is with top running,
> > > the second while the machine should be otherwise completely idle:
> > > 
> > >   ftp://ftp.netbsd.org/pub/NetBSD/misc/simonb/ioflush-gmon.txt.A.bz2
> > >   ftp://ftp.netbsd.org/pub/NetBSD/misc/simonb/ioflush-gmon.txt.B.bz2
> > > 
> > > I've not looked at them myself yet...
> > 
> > Those were very useful, thanks. I have committed another set of fixes.
> > Can you try running with them and see if the problem still occurs?
> 
> After 1 days 20 hours this seems better, but I'm still not sure if the
> "problem" is totally fixed.  Right now top shows ioflush using about 3%
> of CPU time, and it's used about 84 mins since boot:
> 
>   PID   LID USERNAME PRI STATE      TIME   WCPU    CPU COMMAND      NAME
>     0    49 root     124 syncer/0  84:28  3.08%  3.08% [system]     ioflush
> 
> I've only updated my kernel to netbsd-5, but still have old userland.
> I'm tempted to put the old kernel back and see if this amount of CPU
> usage for ioflush is "normal" or not...

Can you try profiling again and maybe add some evcnts in ffs_sync to see
what type of vnode it is flushing (VREG, VDIR, other?).