在 2021/12/15 17:56, zhong jiang 写道:
The patch support multiple target nodes demotion,
but
alloc_migration_target use an gfp_thisnode to
allocate an memory, It maybe fails even though we use an round-robin to
slect the proper demotion node.
Is it possible for this situation ?
Yes, if AEP node is full and failed to allocate memory, then we should
not do memory demotion to AEP, just failed the migration. And now the
kswapd of the AEP node has been waked up to reclaim memory for AEP node
to relieve this issue.
So until now I do not think this is a big problem. In future I think we
can optimize the proactive demotion of DRAM and proactive reclaim of AEP.