On Tue, Jul 03, 2018 at 01:38:59PM -0400, Mathieu Desnoyers wrote: > ----- On Jul 3, 2018, at 1:34 PM, Peter Zijlstra pet...@infradead.org wrote: > > > On Tue, Jul 03, 2018 at 10:10:37AM -0700, Linus Torvalds wrote: > >> On Tue, Jul 3, 2018 at 9:40 AM Andi Kleen <a...@firstfloor.org> wrote: > >> > > >> > So it sounds like architectures that don't have an instruction atomic u64 > >> > *_user need to disable interrupts during the access, and somehow handle > >> > that > >> > case when a page fault happens? > >> > >> No. It's actually the store by *user* space that is the critical one. > >> Not the whole 64-bit value, just the low pointer part. > >> > >> The kernel could do it as a byte-by-byte load, really. It's > >> per-thread, and once the kernel is running, it's not going to change. > >> The kernel never changes the value, it just loads it from user space. > > > > The kernel doesn't change _this_ value, but the kernel does change other > > values, like for instance rseq->cpu_id. But even there, it could use > > byte stores and it is again the userspace load of that field that is > > critical again and needs to be a single op. > > I can simply document that loads/stores from/to all struct rseq fields > should be thread-local then ?
I'm not sure that covers things sufficiently. You really want the userspace load/stores to be single instructions. Also, I think it was rseq_update_cpu_id() where we wanted to use a single u64 store if possible but you worried about the stores.
