On Sun 25-10-20 23:19:34, Matthew Wilcox wrote: > On Thu, Oct 01, 2020 at 09:17:28AM +0200, Jan Kara wrote: > > > I have a followup patch which isn't part of this series which fixes it: > > > > > > http://git.infradead.org/users/willy/pagecache.git/commitdiff/364283163847d1c106463223b858308c730592a1 > > > > Yeah, that looks good. How about partial THPs? The way you've implemented > > it we will now possibly evict more than strictly required. But OTOH > > evicting exactly may require THP split which is a bit unfortunate. But > > probably still a better option because otherwise we could have pages being > > repeatedly brought in and out of cache just because e.g. workload mixes > > direct and buffered IO and is not aligned to THP boundary (and although I > > find loads mixing buffered and direct IO to the same file badly designed, > > I know for a fact that they do exist and if the file ranges are not > > overlapping, it is not that insane design). > > Sorry, forgot to reply to this. > > In this patchset, THPs are created by readahead. We always start > out by allocating order-0 pages and only ramp up after hitting a page > marked as PageReadahead. So it's not like tmpfs where we'll try to jump > straight to order-9 pages and have to worry about the behaviour you're > describing above. That means in this kind of scenario, we might have, > eg, an order-6 page in the cache, remove the whole thing, then bring > back in some order-0 pages. If we hit on those, we'll bring in some > order-2 pages. We won't bring in order-6 pages again until we've hit > in the readahead window twice more. > > I think the ramp-up is probably too aggressive, but it's fun for testing. OK, sounds good then. Thanks for explanation. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
