On Mon, Feb 26, 2024 at 09:17:41PM -0800, Paul E. McKenney wrote: > On Mon, Feb 26, 2024 at 08:08:17PM -0500, Kent Overstreet wrote: > > On Mon, Feb 26, 2024 at 04:55:29PM -0800, Paul E. McKenney wrote: > > > On Mon, Feb 26, 2024 at 07:29:04PM -0500, Kent Overstreet wrote: > > > > On Mon, Feb 26, 2024 at 04:05:37PM -0800, Paul E. McKenney wrote: > > > > > On Mon, Feb 26, 2024 at 06:29:43PM -0500, Kent Overstreet wrote: > > > > > > Well, we won't want it getting hammered on continuously - we should be > > > > > > able to tune reclaim so that doesn't happen. > > > > > > > > > > > > I think getting numbers on the amount of memory stranded waiting for RCU > > > > > > is probably first order of business - minor tweak to kfree_rcu() et all > > > > > > for that; there's APIs they can query to maintain that counter. > > > > > > > > > > We can easily tell you the number of blocks of memory waiting to be freed. > > > > > But RCU does not know their size. Yes, we could ferret this on each > > > > > call to kmem_free_rcu(), but that might not be great for performance. > > > > > We could traverse the lists at runtime, but such traversal must be done > > > > > with interrupts disabled, which is also not great. > > > > > > > > > > > then, we can add a heuristic threshhold somewhere, something like > > > > > > > > > > > > if (rcu_stranded * multiplier > reclaimable_memory) > > > > > > kick_rcu() > > > > > > > > > > If it is a heuristic anyway, it sounds best to base the heuristic on > > > > > the number of objects rather than their aggregate size. > > > > > > > > I don't think that'll really work given that object size can very from < > > > > 100 bytes all the way up to 2MB hugepages. The shrinker API works that > > > > way and I positively hate it; it's really helpful for introspection and > > > > debugability later to give good human understandable units to this > > > > stuff. > > > > > > You might well be right, but let's please try it before adding overhead to > > > kfree_rcu() and friends. I bet it will prove to be good and sufficient. > > > > > > > And __ksize() is pretty cheap, and I think there might be room in struct > > > > slab to stick the object size there instead of getting it from the slab > > > > cache - and folio_size() is cheaper still. > > > > > > On __ksize(): > > > > > > * This should only be used internally to query the true size of allocations. > > > * It is not meant to be a way to discover the usable size of an allocation > > > * after the fact. Instead, use kmalloc_size_roundup(). > > > > > > Except that kmalloc_size_roundup() doesn't look like it is meant for > > > this use case. On __ksize() being used only internally, I would not be > > > at all averse to kfree_rcu() and friends moving to mm. > > > > __ksize() is the right helper to use for this; ksize() is "how much > > usable memory", __ksize() is "how much does this occupy". > > > > > The idea is for kfree_rcu() to invoke __ksize() when given slab memory > > > and folio_size() when given vmalloc() memory? > > > > __ksize() for slab memory, but folio_size() would be for page > > allocations - actually, I think compound_order() is more appropriate > > here, but that's willy's area. IOW, for free_pages_rcu(), which AFAIK we > > don't have yet but it looks like we're going to need. > > > > I'm scanning through vmalloc.c and I don't think we have a helper yet to > > query the allocation size - I can write one tomorrow, giving my brain a > > rest today :) > > Again, let's give the straight count of blocks a try first. I do see > that you feel that the added overhead is negligible, but zero added > overhead is even better. How are you going to write a heuristic that works correctly both when the system is cycling through nothing but 2M hugepages, and nothing but 128 byte whatevers?
