On Thu, Oct 22, 2020 at 12:04:22AM +0100, Matthew Wilcox wrote: > On Thu, Oct 22, 2020 at 09:14:35AM +1100, Dave Chinner wrote: > > On Tue, Oct 20, 2020 at 11:53:31PM +0100, Matthew Wilcox wrote: > > > True, we don't _have to_ split THP on holepunch/truncation/... but it's > > > a better implementation to free pages which cover blocks that no longer > > > have data associated with them. > > > > "Better" is a very subjective measure. What numbers do you have > > to back that up? > > None. When we choose to use a THP, we're choosing to treat a chunk > of a file as a single unit for the purposes of tracking dirtiness, > age, membership of the workingset, etc. We're trading off reduced > precision for reduced overhead; just like the CPU tracks dirtiness on > a cacheline basis instead of at byte level. > > So at some level, we've making the assumption that this 128kB THP is > all one thingand it should be tracked together. But the user has just > punched a hole in it. I can think of no stronger signal to say "The > piece before this hole, the piece I just got rid of and the piece after > this are three separate pieces of the file". There's a difference between the physical layout of the file and representing data efficiently in the page cache. Just because we can use a THP to represent a single extent doesn't mean we should always use that relationship, nor should we require that small manipulations of on-disk extent state require that page cache pages be split or gathered. i.e. the whole point of the page cache is to decouple the physical layout of the file from the user access mechanisms for performance reasons, not tie them tightly together. I think that's the wrong approach to be taking here - truncate/holepunch do not imply that THPs need to be split unconditionally. Indeed, readahead doesn't care that a THP might be split across mulitple extents and require multiple bios to bring tha data into cache, so why should truncate/holepunch type operations require the THP to be split to reflect underlying disk layouts? > If I could split them into pieces that weren't single pages, I would. > Zi Yan has a patch to do just that, and I'm very much looking forward > to that being merged. But saying "Oh, this is quite small, I'll keep > the rest of the THP together" is conceptually wrong. Yet that's exactly what we do with block size < PAGE_SIZE configurations, so I fail to see why it's conceptually wrong for THPs to behave the same way and normal pages.... > > > Splitting the page instead of throwing it away makes sense once we can > > > transfer the Uptodate bits to each subpage. If we don't have that, > > > it doesn't really matter which we do. > > > > Sounds like more required functionality... > > I'm not saying that my patchset is the last word and there will be no > tweaking. I'm saying I think it's good enough, an improvement on the > status quo, and it's better to merge it for 5.11 than to keep it out of > tree for another three months while we tinker with improving it. > > Do you disagree? In part. Concepts and algorithms need to be sound and agreed upon before we merge patches, and right now I disagree with the some of the basic assumptions about how THP and filesystem layout operations are being coupled. That part needs to be sorted before stuff gets merged... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
