Hi. On Fri, May 13, 2022 at 11:54:16AM -0700, Roman Gushchin <roman.gushchin@xxxxxxxxx> wrote: > Hm, what are our plans here? Are we going to fix it soon-ish, or there > is still no agreement on how to proceed? Here are some of my ideas in random order so far and comments: 0) mask memory.events:low -> not a real fix 1) don't do SWAP_CLUSTER_MAX roundup -> won't solve sudden lift of protection 2) instead of SWAP_CLUSTER_MAX over-reclaim, do same under-reclaim -> same problem as 1) 3) update children_low_usage transactionally (after reclaim round is done) - ??? 4) don't recursively distribute residual protection in the same reclaim round - ??? 5) iterate siblings from highest to lowest protection - not a solution 6) assign only >SWAP_CLUSTER_MAX of residuum - need more info I'm discouraged by possible complexity of 3) or 4), while 6) is what I'd like to look more into. HTH, Michal
