On Fri, Mar 14, 2025 at 12:57:39PM -0400, Johannes Weiner wrote: > On Fri, Mar 14, 2025 at 03:49:30PM +0100, Michal Hocko wrote: > > On Fri 14-03-25 10:18:33, Johannes Weiner wrote: > > > On Fri, Mar 14, 2025 at 10:27:57AM +0100, Michal Hocko wrote: > > [...] > > > > I have just noticed that you have followed up [1] with a concern that > > > > using swappiness in the whole min-max range without any heuristics turns > > > > out to be harder than just relying on the min and max as extremes. > > > > What seems to be still missing (or maybe it is just me not seeing that) > > > > is why should we only enforce those extreme ends of the range and still > > > > preserve under-defined semantic for all other swappiness values in the > > > > pro-active reclaim. > > > > > > I'm guess I'm not seeing the "under-defined" part. > > > > What I meant here is that any other value than both ends of swappiness > > doesn't have generally predictable behavior unless you know specific > > details of the current memory reclaim heuristics in get_scan_count. > > > > > cache_trim_mode is > > > there to make sure a streaming file access pattern doesn't cause > > > swapping. > > > > Yes, I am aware of the purpose. > > > > > He has a special usecase to override cache_trim_mode when he > > > knows a large amount of anon is going cold. There is no way we can > > > generally remove it from proactive reclaim. > > > > I believe I do understand the requirement here. The patch offers > > counterpart to noswap pro-active reclaim and I do not have objections to > > that. > > > > The reason I brought this up is that everything in between 0..200 is > > kinda gray area. We've had several queries why swappiness=N doesn't work > > as expected and the usual answer was because of heuristics. Most people > > just learned to live with that and stopped fine tuning vm_swappiness. > > Which is good I guess. > > You're still oversimplifying and then dismissing. The heuristics don't > make swappiness meaningless, they make it useful in the first place. > > This control is used to define the rough relative IO cost of swapping > and filesystem paging, as a value between 0 and 200. > > This is clearly defined, and implemented as such. cache_trim_mode is > predicated on the *absence* of paging and caching benefits: A linear, > use-once file access pattern that *does not* benefit from additional > cache space. Kicking out anon for that purpose would be wrong under > pretty much any circumstance. That's why it "overrides" swappiness: > swappiness cannot apply when swapping at all would be nonsense. > > Proactive reclaimers like ours rely on this. We use swappiness to > express exactly what it says on the tin: the relative cost between > thrashing file vs anon. We use it quite effectively to manage anon > write rates for flash wear management e.g. Obviously that doesn't mean > we want to swap when somebody streams through a large file set. > > Zhongkun's case is a significant exception. He just wants to get rid > of known-cold anon set. This level of insight into userspace access > patterns is rare in practice. You could argue that MADV_PAGEOUT might > be more suitable for that. We have a similar use case at Google where we have a known-cold anon set and we proactively reclaim it. This is why we previously proposed type=anon/file/.., but swappiness is more flexible for use cases like the one Johannes describes above. > But I also don't necessarily see a problem > with making swappiness=200 do it; although we might have to teach our > proactive reclaimer to auto-tune between 1 and 199 then. Would it be better if we don't use the existing swappiness=200 for this? We can support something like memory.reclaim X swappiness=max instead to achieve the "anon only" mode without affecting the existing semantics of swappiness at all. I have a feeling I may have already proposed that at some point. In the kernel, we can define a new value (say 201 or 1000) that means anon only and set it in memory_reclaim() when "max" is specified. We can then explicitly check for this value in get_scan_count() (we probably also need to handle MGLRU?).