Re: + mm-vmscan-add-mm_vmscan_inactive_list_is_low-tracepoint.patch added to -mm tree

[Michal Hocko]

On Thu 12-01-17 17:48:13, Minchan Kim wrote:
> On Thu, Jan 12, 2017 at 09:15:54AM +0100, Michal Hocko wrote:
> > On Thu 12-01-17 14:12:47, Minchan Kim wrote:
> > > Hello,
> > > 
> > > On Wed, Jan 11, 2017 at 04:52:39PM +0100, Michal Hocko wrote:
> > > > On Wed 11-01-17 08:52:50, Minchan Kim wrote:
> > > > [...]
> > > > > > @@ -2055,8 +2055,8 @@ static bool inactive_list_is_low(struct
> > > > > >  	if (!file && !total_swap_pages)
> > > > > >  		return false;
> > > > > >  
> > > > > > -	inactive = lruvec_lru_size(lruvec, file * LRU_FILE);
> > > > > > -	active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE);
> > > > > > +	total_inactive = inactive = lruvec_lru_size(lruvec, file * LRU_FILE);
> > > > > > +	total_active = active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE);
> > > > > >  
> > > > > 
> > > > > the decision of deactivating is based on eligible zone's LRU size,
> > > > > not whole zone so why should we need to get a trace of all zones's LRU?
> > > > 
> > > > Strictly speaking, the total_ counters are not necessary for making the
> > > > decision. I found reporting those numbers useful regardless because this
> > > > will give us also an information how large is the eligible portion of
> > > > the LRU list. We do not have any other tracepoint which would report
> > > > that.
> > > 
> > > The patch doesn't say anything why it's useful. Could you tell why it's
> > > useful and inactive_list_is_low should be right place?
> > > 
> > > Don't get me wrong, please. I don't want to bother you.
> > > I really don't want to add random stuff although it's tracepoint for
> > > debugging.
> > 
> > This doesn't sounds random to me. We simply do not have a full picture
> > on 32b systems without this information. Especially when memcgs are
> > involved and global numbers spread over different LRUs.
> 
> Could you elaborate it?

The problem with 32b systems is that you only can consider a part of the
LRU for the lowmem requests. While we have global counters to see how
much lowmem inactive/active pages we have, those get distributed to
memcg LRUs. And that distribution is impossible to guess. So my thinking
is that it can become a real head scratcher to realize why certain
active LRUs are aged while others are not. This was the case when I was
debugging the last issue which triggered all this. All of the sudden I
have seen many invocations when inactive and active were zero which
sounded weird, until I realized that those are memcg's lruvec which is
what total numbers told me...

Later on I would like to add an memcg identifier to the vmscan
tracepoints but I didn't get there yet.
 
> "
> Currently we have tracepoints for both active and inactive LRU lists
> reclaim but we do not have any which would tell us why we we decided to
> age the active list.  Without that it is quite hard to diagnose
> active/inactive lists balancing.  Add mm_vmscan_inactive_list_is_low
> tracepoint to tell us this information.
> "
> 
> Your description says "why we decided to age the active list".
> So, what's needed?
> 
> >  
> > > > [...]
> > > > > > @@ -2223,7 +2228,7 @@ static void get_scan_count(struct lruvec
> > > > > >  	 * lruvec even if it has plenty of old anonymous pages unless the
> > > > > >  	 * system is under heavy pressure.
> > > > > >  	 */
> > > > > > -	if (!inactive_list_is_low(lruvec, true, sc) &&
> > > > > > +	if (!inactive_list_is_low(lruvec, true, sc, false) &&
> > > > > 
> > > > > Hmm, I was curious why you added trace boolean arguement and found it here.
> > > > > Yes, here is not related to deactivation directly but couldn't we help to
> > > > > trace it unconditionally?
> > > > 
> > > > I've had it like that when I was debugging the mentioned bug and found
> > > > it a bit disturbing. It generated more output than I would like and it
> > > > wasn't really clear from which code path  this has been called from.
> > > 
> > > Indeed.
> > > 
> > > Personally, I want to move inactive_list_is_low in shrink_active_list
> > > and shrink_active_list calls inactive_list_is_low(...., true),
> > > unconditionally so that it can make code simple/clear but cannot remove
> > > trace boolean variable , which what I want. So, it's okay if you love
> > > your version.
> > 
> > I am not sure I am following. Why is the additional parameter a problem?
> 
> Well, to me, it's not a elegance. Is it? If we need such boolean variable
> to control show the trace, it means it's not a good place or think
> refactoring.

But, even when you refactor the code there will be other callers of
inactive_list_is_low outside of shrink_active_list...

-- 
Michal Hocko
SUSE Labs