Re: [PATCH v3] mm: migrate: Support multiple target nodes demotion

[Baolin Wang]

On 2021/11/12 11:02, Huang, Ying wrote:
> Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx> writes:
>
> > On 2021/11/12 10:44, Huang, Ying wrote:
> > > Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx> writes:
> > >
> > > > We have some machines with multiple memory types like below, which
> > > > have one fast (DRAM) memory node and two slow (persistent memory) memory
> > > > nodes. According to current node demotion policy, if node 0 fills up,
> > > > its memory should be migrated to node 1, when node 1 fills up, its
> > > > memory will be migrated to node 2: node 0 -> node 1 -> node 2 ->stop.
> > > >
> > > > But this is not efficient and suitbale memory migration route
> > > > for our machine with multiple slow memory nodes. Since the distance
> > > > between node 0 to node 1 and node 0 to node 2 is equal, and memory
> > > > migration between slow memory nodes will increase persistent memory
> > > > bandwidth greatly, which will hurt the whole system's performance.
> > > >
> > > > Thus for this case, we can treat the slow memory node 1 and node 2
> > > > as a whole slow memory region, and we should migrate memory from
> > > > node 0 to node 1 and node 2 if node 0 fills up.
> > > >
> > > > This patch changes the node_demotion data structure to support multiple
> > > > target nodes, and establishes the migration path to support multiple
> > > > target nodes with validating if the node distance is the best or not.
> > > >
> > > > available: 3 nodes (0-2)
> > > > node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
> > > > node 0 size: 62153 MB
> > > > node 0 free: 55135 MB
> > > > node 1 cpus:
> > > > node 1 size: 127007 MB
> > > > node 1 free: 126930 MB
> > > > node 2 cpus:
> > > > node 2 size: 126968 MB
> > > > node 2 free: 126878 MB
> > > > node distances:
> > > > node   0   1   2
> > > >     0:  10  20  20
> > > >     1:  20  10  20
> > > >     2:  20  20  10
> > > >
> > > > Signed-off-by: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>
> >
> > snip
> >
> > > >    	/*
> > > >    	 * 'next_pass' contains nodes which became migration
> > > > @@ -3192,6 +3281,14 @@ static int __init migrate_on_reclaim_init(void)
> > > >    {
> > > >    	int ret;
> > > >    +	/*
> > > > +	 * Ignore allocation failure, if this kmalloc fails
> > > > +	 * at boot time, we are likely in bigger trouble.
> > > > +	 */
> > > > +	node_demotion = kmalloc_array(nr_node_ids,
> > > > +				      sizeof(struct demotion_nodes),
> > > > +				      GFP_KERNEL);
> > > > +
> > > I think we should WARN_ON() here.
> >
> > In this unlikey case, I think the mm core will print more information,
> > IMHO WARN_ON() will help little. Anyway no strong opinion on
> > this. Other than that, can I get your reviewed-by tag with this nit
> > fixed? Thanks.
>
> Yes.  Please add my "reviewed-by" after changing this.

OK. Thanks for your reviewing.