On Thu, Oct 24, 2019 at 4:40 PM Andrea Parri <parri.andrea@xxxxxxxxx> wrote: > > On Thu, Oct 24, 2019 at 03:58:40PM +0200, Dmitry Vyukov wrote: > > On Thu, Oct 24, 2019 at 3:43 PM Andrea Parri <parri.andrea@xxxxxxxxx> wrote: > > > > > > > But why? I think kernel contains lots of such cases and it seems to be > > > > officially documented by the LKMM: > > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/memory-model/Documentation/explanation.txt > > > > address dependencies and ppo > > > > > > Well, that same documentation also alerts about some of the pitfalls > > > developers can incur while relying on dependencies. I'm sure you're > > > more than aware of some of the debate surrounding these issues. > > > > I thought that LKMM is finally supposed to stop all these > > centi-threads around subtle details of ordering. And not we finally > > have it. And it says that using address-dependencies is legal. And you > > are one of the authors. And now you are arguing here that we better > > not use it :) Can we have some black/white yes/no for code correctness > > reflected in LKMM please :) If we are banning address dependencies, > > don't we need to fix all of rcu uses? > > Current limitations of the LKMM are listed in tools/memory-model/README > (and I myself discussed a number of them at LPC recently); the relevant > point here seems to be: > > 1. Compiler optimizations are not accurately modeled. Of course, > the use of READ_ONCE() and WRITE_ONCE() limits the compiler's > ability to optimize, but under some circumstances it is possible > for the compiler to undermine the memory model. [...] > > Note that this limitation in turn limits LKMM's ability to > accurately model address, control, and data dependencies. > > A less elegant, but hopefully more effective, way to phrase such point > is maybe "feel free to rely on dependencies, but then do not blame the > LKMM authors please". ;-) We are not going to blame LKMM authors :) Acquire will introduce actual hardware barrier on arm/power/etc. Maybe it does not matter here. But I feel if we start replacing all load-depends/rcu with acquire, it will be noticeable overhead. So what do we do in the context of the whole kernel?