Yeah, I aware what you're talking about. In fact most kernels has common/shared page-cache (since it is just reasonable and useful), so therefore such hack is works. Nevertheless it would be great to make msync(MS_INVALIDATE) work for this case, i.e. for read-only mmaps from the fd-writer side.
чт, 14 нояб. 2019 г. в 18:02, Janne Johansson <[email protected]>: > > I'm aware that many/most OSes do sync it, but > > Den tors 14 nov. 2019 kl 15:08 skrev Леонид Юрьев <[email protected]>: >> >> Thank you for the clarification. >> "It's Not a Bug, It's a Feature" ;) > > Something like that. > It is one of those things that could be neat if it was, but seemingly > non-trivial to make happen, and there is no promise anywhere that any and all > kinds of accesses to some data is synched for you, even if most other OSes > does that for you nowadays. > >>> That is how OpenBSD works, you have to decide if you want to talk mmap() or >>> via file descriptor, not both at the same time.
