On Fri, 13 Mar 2020, David Rientjes wrote: > I'll renew my suggestion of defaulting memcg->swappiness to -1 and then > letting mem_cgroup_swappiness() return vm_swappiness when > memcg->swappiness == -1. > > I don't think the act of creating a memcg and not initializing the > swappiness value implies that the existing value meets your expectation. > Thanks, David, I haven't considered this point. This point has made me realising that the existing behaviour, where the swappiness is inherited from the parent, is not documented and, apparently, just appears to be implemented that way. In this case, the -1 approach looks less harmful in comparison with the original patch, unless I underestimate a number of users "affected" by the assumptions about how memcg->swappiness is getting propagated. I'll send another patch implementing the -1 approach shortly, perhaps with some clarifications in the "swappiness" documentation.
