On 2020-01-15 23:04:19 [+0100], Vlastimil Babka wrote:

On 1/15/2020 5:10 PM, Sebastian Andrzej Siewior wrote:

quoted

Since commit
    5bbe3547aa3ba ("mm: allow compaction of unevictable pages")

it is allowed to examine mlocked pages and compact them by default.
On -RT even minor pagefaults are problematic because it may take a few
100us to resolve them and until then the task is blocked.

Fine, this makes sense on RT I guess. There might be some trade-off for
high-order allocation latencies though. We could perhaps migrate such mlocked
pages to pages allocated without __GFP_MOVABLE during the mlock() to at least
somewhat prevent them being scattered all over the zones. For MCL_FUTURE,
allocate them as unmovable from the beginning. But that can wait until issues
are reported.
I assume you have similar solution for NUMA balancing and whatever else can
cause minor faults?

I've found this one while testing. Could you please point to the NUMA
balancing that might be an issue?

quoted

Make compact_unevictable_allowed = 0 default and remove it from /proc on
RT.

Removing it is maybe going too far in terms of RT kernel differences confusing
users? Change the default sure, perhaps making it read-only, but removing?

Okay. I will make it RO then. 

quoted

Link: https://lore.kernel.org/linux-mm/20190710144138.qyn4tuttdq6h7kqx@linutronix.de/ (local)

In any case the sysctl Documentation/ should be updated? And perhaps also the
mlock manpage as you noted in the older thread above?

Sure. Let me add the sysctl documentation to this patch and then I will
look into the manpage.

Thanks,
Vlastimil

Sebastian