On Sun, Feb 17, 2019 at 01:33:36PM -0800, Davidlohr Bueso wrote: > On Fri, 15 Feb 2019, Jason Gunthorpe wrote: > > This looks just wrong... The algorithm should look like the one in umem > > On another note, and considering infiniband (core) is allowed on most > archs, do we want to relax the ordering guarantees from the compound > atomic64_inc_return(pinned_vm) in ib_umem_get()? > Looking at IB core and some of the drivers I don't see any particular > constraints on the order the pinned_vm is touched - so ll/sc based archs > could benefit with atomic64_inc_return_relaxed(). I don't think order matters for any place touching pinned_vm, but it also doesn't seem like any of these are such a performance path the difference matters Jason
