Re: [v5] mm: khugepaged: recalculate min_free_kbytes after memory hotplug as expected by khugepaged

[Michal Hocko]

On Wed 30-09-20 15:03:11, Mike Kravetz wrote:
> On 9/30/20 1:47 PM, Vijay Balakrishna wrote:
> > On 9/30/2020 11:20 AM, Mike Kravetz wrote:
> >> On 9/29/20 9:49 AM, Vijay Balakrishna wrote:
> >>
> >> Sorry for jumping in so late.  Should we use this as an opportunity to
> >> also fix up the messages logged when (re)calculating mfk?  They are wrong
> >> and could be quite confusing.
> > 
> > 
> > Sure.  Please share your thoughts regarding appropriate message.  Here is what I'm thinking
> > 
> > pr_warn("min_free_kbytes is not updated to %d because current value %d is preferred\n", new_min_free_kbytes, min_free_kbytes);
> > 
> > If above message is reasonable I can post a new revision (v6).
> 
> Just considering the below example,
> 
> >> For example consider the following sequence
> >> of operations and corresponding log messages produced.
> >>
> >> Freshly booted VM with 2 nodes and 8GB memory:
> >> # cat /proc/sys/vm/min_free_kbytes
> >> 90112
> >> # echo 90000 > /proc/sys/vm/min_free_kbytes
> >> # cat /proc/sys/vm/min_free_kbytes
> >> 90000
> >> # echo 0 > /sys/devices/system/node/node1/memory56/online
> >> [  135.099947] Offlined Pages 32768
> >> [  135.102362] min_free_kbytes is not updated to 11241 because user defined value 90000 is preferred
> 
> I am not sure if there is any value in printing the above line.  Especially
> in this context as it becomes obsolete with the printing of the next line.

The original intention was to make it explicit that auto-tuning is
influenced by the user provided configuration.

> >> [  135.109070] khugepaged: raising min_free_kbytes from 90000 to 90112 to help t
> >> ransparent hugepage allocations
> 
> IMO, the above line is the only one that should be output as a result of the
> recalculation.

Well, but khugepaged could be disabled and then the above might not get
printed. Sure the code could get reorganized and all that but is this
really worth that?

> I guess that brings up the question of 'should we continue to track the user
> defined value if we overwrite it?".  If we quit tracking it may help with the
> next message.

Auto tuning and user provided override is quite tricky to get sensible.
Especially in the case here. Admin has provided an override but has the
potential memory hotplug been considered? Or to make it even more
complicated, consider that the hotplug happens without admin involvement
- e.g. memory gets hotremoved due to HW problems. Is the admin provided
value still meaningful? To be honest I do not have a good answer and I
am not sure we should care all that much until we see practical
problems.
-- 
Michal Hocko
SUSE Labs