On Mon, Sep 08, 2025 at 10:35:38AM -0300, Jason Gunthorpe wrote: > On Mon, Sep 08, 2025 at 02:27:12PM +0100, Lorenzo Stoakes wrote: > > > It's not only remap that is a concern here, people do all kinds of weird > > and wonderful things in .mmap(), sometimes in combination with remap. > > So it should really not be split this way, complete is a badly name I don't understand, you think we can avoid splitting this in two? If so, I disagree. We have two stages, _intentionally_ placed to avoid the issues the mmap_prepare series in the first instance worked to avoid: 1. 'Hey, how do we configure this VMA we have _not yet set up_' 2. 'OK it's set up, now do you want to do something else? I'm sorry but I'm not sure how we could otherwise do this. Keep in mind re: point 1, we _need_ the VMA to be established enough to check for merge etc. Another key aim of this change was to eliminate the need for a merge re-check. > prepopulate and it should only fill the PTEs, which shouldn't need > more locking. > > The only example in this series didn't actually need to hold the lock. There's ~250 more mmap callbacks to work through. Do you provide a guarantee that: - All 250 absolutely only need access to the VMAs to perform prepopulation of this nature. - That absolutely none will set up state in the prepopulate step that might need to be unwound should an error arise? Keeping in mind I must remain practical re: refactoring each caller. I mean, let me go check what you say re: the resctl lock, if you're right I could drop mmap_abort for now and add it later if needed. But re: calling mmap_complete prepopulate, I don't really think that's sensible. mmap_prepare is invoked at the point of the preparation of the mapping, and mmap_complete is invoked once that preoparation is complete to allow further actions. I'm obviously open to naming suggestions, but I think it's safer to consistently refer to where we are in the lifecycle rather than presuming what the caller might do. (I'd _prefer_ they always did just prepopulate, but I just don't think we necessarily can). > > Jason Cheers, Lorenzo
