On Thu 17-09-20 19:09:00, Jason Gunthorpe wrote: > On Thu, Sep 17, 2020 at 05:40:59PM -0400, Peter Xu wrote: > > On Thu, Sep 17, 2020 at 01:35:56PM -0700, Linus Torvalds wrote: > > > For that to happen, we'd need to have the vma flag so that we wouldn't > > > have any worry about non-pinners, but as you suggested, I think even > > > just a mm-wide counter - or flag - to deal with the fast-bup case is > > > likely perfectly sufficient. > > > > Would mm_struct.pinned_vm suffice? > > I think that could be a good long term goal > > IIRC last time we dug into the locked_vm vs pinned_vm mess it didn't > get fixed. There is a mix of both kinds, as you saw, and some > resistance I don't clearly remember to changing it. > > My advice for this -rc fix is to go with a single bit in the mm_struct > set on any call to pin_user_pages* > > Then only users using pin_user_pages and forking are the only ones who > would ever do extra COW on fork. I think that is OK for -rc, this > workload should be rare due to the various historical issues. Anyhow, > a slow down regression is better than a it is broken regression. Agreed. I really like the solution of not write-protecting pinned pages on fork(2). > This can be improved into a counter later. Due to the pinned_vm > accounting all call sites should have the mm_struct at unpin, but I > have a feeling it will take a alot of driver patches to sort it all > out. I somewhat fear that some of the users of pin_user_pages() don't bother with pinned_vm accounting exactly because they don't have mm_struct on unpin... Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
