On Wed, Sep 04, 2019 at 03:11:10PM +0100, Will Deacon wrote: > Hi Joel, > > On Wed, Sep 04, 2019 at 09:24:18AM -0400, Joel Fernandes wrote: > > On Wed, Sep 04, 2019 at 01:30:38PM +0200, Peter Zijlstra wrote: > > > On Wed, Sep 04, 2019 at 06:16:16AM -0400, Steven Rostedt wrote: > > > > On Mon, 2 Sep 2019 11:16:23 +0200 > > > > Peter Zijlstra <pet...@infradead.org> wrote: > > > > > > > > > in sched_dl_period_handler(). And do: > > > > > > > > > > + preempt_disable(); > > > > > max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * > > > > > NSEC_PER_USEC; > > > > > min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * > > > > > NSEC_PER_USEC; > > > > > + preempt_enable(); > > > > > > > > Hmm, I'm curious. Doesn't the preempt_disable/enable() also add > > > > compiler barriers which would remove the need for the READ_ONCE()s here? > > > > > > They do add compiler barriers; but they do not avoid the compiler > > > tearing stuff up. > > > > Neither does WRITE_ONCE() on some possibly buggy but currently circulating > > compilers :( > > Hmm. The example above is using READ_ONCE, which is a different kettle of > frogs.
True. But, I equally worry about all *-tearing frog kettles ;-) > > As Will said in: > > https://lore.kernel.org/lkml/20190821103200.kpufwtviqhpbuv2n@willie-the-truck/ > > > > void bar(u64 *x) > > { > > *(volatile u64 *)x = 0xabcdef10abcdef10; > > } > > > > gives: > > > > bar: > > mov w1, 61200 > > movk w1, 0xabcd, lsl 16 > > str w1, [x0] > > str w1, [x0, 4] > > ret > > > > Speaking of which, Will, is there a plan to have compiler folks address this > > tearing issue and are bugs filed somewhere? I believe aarch64 gcc is buggy, > > and clang is better but is still buggy? > > Well, it depends on your point of view. Personally, I think tearing a > volatile access (e.g. WRITE_ONCE) is buggy and it seems as though the GCC > developers agree: > > https://gcc.gnu.org/ml/gcc-patches/2019-08/msg01500.html > > so it's likely this will be fixed for AArch64 GCC. I couldn't persuade > clang to break the volatile case, so think we're good there too. Glad to know that GCC folks are looking into the issue. Sorry if this is getting a bit off-topic. Also does the aarch64 clang doing the "memset folding" issue, also need to be looked into? You had mentioned it in the same thread: https://lore.kernel.org/lkml/20190821103200.kpufwtviqhpbuv2n@willie-the-truck/ Or, does WRITE_ONCE() resolve such memset store-merging? > For the non-volatile case, I don't actually consider it to be a bug, > although I sympathise with the desire to avoid a retrospective tree-wide > sweep adding random WRITE_ONCE invocations to stores that look like they > might be concurrent. In other words, I think I'd suggest: > > * The use of WRITE_ONCE in new code (probably with a comment justifying it) > * The introduction of WRITE_ONCE to existing code where it can be shown to > be fixing a real bug (e.g. by demonstrating that a compiler actually > gets it wrong) > > For the /vast/ majority of cases, the compiler will do the right thing > even without WRITE_ONCE, simply because that's going to be the most > performant choice as well. Thanks for the thoughts. They seem to be reasonable to me. thanks, - Joel
