On 09.08.22 21:07, Linus Torvalds wrote: > On Tue, Aug 9, 2022 at 11:48 AM Jason Gunthorpe <jgg@xxxxxxxxxx> wrote: >> >> It is because of all this madness with COW. > > Yes, yes, but we have the proper long-term pinning now with > PG_anon_exclusive, and it actually gets the pinning right not just > over COW, but even over a fork - which that early write never did. > > David, I thought all of that got properly merged? Is there something > still missing? The only thing to get R/O longterm pins in MAP_PRIVATE correct that's missing is that we have to break COW when taking a R/O longterm pin when *not* finding an anon page inside a private mapping. Regarding anon pages I am not aware of issues (due to PG_anon_exclusive). If anybody here wants to stare at a webpage, the following commit explains the rough idea for MAP_PRIVATE: https://github.com/davidhildenbrand/linux/commit/cd7989fb76d2513c86f01e6f7a74415eee5d3150 Once we have that in place, we can mostly get rid of FOLL_FORCE|FOLL_WRITE for R/O longterm pins. There are some corner cases though that need some additional thought which i am still working on. FS-handled COW in MAP_SHARED mappings is just nasty (hello DAX). (the wrong use of FOLL_GET instead of FOLL_PIN for O_DIRECT and friends still persists, but that's a different thing to handle and it's only problematic with concurrent fork() IIRC) -- Thanks, David / dhildenb
