Re: xen-blkback: Scheduled work from previous purge is still busy, cannot purge list

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Thu, Oct 15, 2020 at 02:53:34PM +0200, J. Roeleveld wrote:
> On Thursday, October 15, 2020 2:00:46 PM CEST Roger Pau Monné wrote:
> > Please don't drop xen-devel mailing list when replying.
> 
> My apologies, most mailing lists I am active on have a working "reply" 
> button. 
> Here I need to use "reply-all".
> 
> 
> > On Thu, Oct 15, 2020 at 01:28:49PM +0200, J. Roeleveld wrote:
> > > On Thursday, October 15, 2020 12:57:35 PM CEST you wrote:
> > > > On Tue, Oct 13, 2020 at 07:26:47AM +0200, J. Roeleveld wrote:
> > > > > Hi All,
> > > > > 
> > > > > I am seeing the following message in the "dmesg" output of a driver
> > > > > domain.
> > > > > 
> > > > > [Thu Oct  8 20:57:04 2020] xen-blkback: Scheduled work from previous
> > > > > purge
> > > > > is still busy, cannot purge list
> > > > > [Thu Oct  8 20:57:11 2020] xen-blkback: Scheduled work from previous
> > > > > purge
> > > > > is still busy, cannot purge list
> > > > > [Thu Oct  8 20:57:44 2020] xen-blkback: Scheduled work from previous
> > > > > purge
> > > > > is still busy, cannot purge list
> > > > > [Thu Oct  8 20:57:44 2020] xen-blkback: Scheduled work from previous
> > > > > purge
> > > > > is still busy, cannot purge list
> > > > > 
> > > > > 
> > > > > Is this something to worry about? Or can I safely ignore this?
> > > > 
> > > > What version of the Linux kernel are you running in that driver
> > > > domain?
> > > 
> > > Host:
> > > Kernel: 5.4.66
> > > Xen: 4.12.3
> > > 
> > > Driver domain:
> > > Kernel: 5.4.66
> > > Xen: 4.12.3
> > > 
> > > > Is the domain very busy? That might explain the delay in purging
> > > > grants.
> > > 
> > > No, it's generally asleep, been going through the munin-records and can't
> > > find any spikes the correlate with the messages either.
> > > 
> > > > Also is this an sporadic message, or it's constantly repeating?
> > > 
> > > It's sporadic, but occasionally, I get it several times in a row.
> > > 
> > > My understanding of the code where this message comes from is far from
> > > sufficient. Which means I have no clue what it is actually trying to do.
> > 
> > There's a recurrent worker thread in blkback that will go and purge
> > unused cache entries after they have expired. This is done to prevent
> > the cache from growing unbounded.
> > 
> > AFAICT this just means the worker is likely running faster than what
> > you can proceed, and hence you get another worker run before the old
> > entries have been removed. Should be safe to ignore, but makes me
> > wonder if I should add a parameter to tune the periodicity of the
> > purge work.
> 
> In other words, when it "fails" in this manner, the queue will simply be left 
> and processed the next time?

Yes, exactly.

> How often does this currently run?

The purge worked will run every 100ms, and the queued work should be
terminated before the next run.

> A parameter to tune the periodicity might be an option, for now I feel 
> confident I can safely ignore these messages.

Sure, I'm testing a patch series to that effect now.

Thanks, Roger.