On Thu, Dec 12, 2019 at 05:04:27PM +0000, Will Deacon wrote:

On Thu, Dec 12, 2019 at 11:46:10AM +0100, Peter Zijlstra wrote:

quoted

On Thu, Dec 12, 2019 at 10:07:56AM +0000, Will Deacon wrote:

quoted

quoted

So your proposed change _should_ be fine. Will, I'm assuming you never
saw this on your ARGH64 builds when you did this code ?

I did see it, but (a) looking at the code out-of-line makes it look a lot
worse than it actually is (so the ext4 example is really helpful -- thanks
Michael!) and (b) I chalked it up to a crappy compiler.

However, see this comment from Arnd on my READ_ONCE series from the other
day:

https://lore.kernel.org/lkml/CAK8P3a0f=WvSQSBQ4t0FmEkcFE_mC3oARxaeTviTSkSa-D2qhg@mail.gmail.com (local)

In which case, I'm thinking that we should be doing better in READ_ONCE()
for non-buggy compilers which would also keep the KCSAN folks happy for this
code (and would help with [1] too).

So something like this then? Although I suppose that should be moved
into compiler-gcc.h and then guarded by #ifndef READ_ONCE or so.

Ah wait, I think we've been looking at this wrong. The volatile pointer
argument is actually the problem here, not READ_ONCE()! The use of typeof()
means that the temporary variable to which __READ_ONCE_SIZE writes ends up
being a volatile store, so it can't be optimised away. This is why we get
a stack access and why stack protector then wrecks the codegen for us.

Hmm, it's actually probably the volatile read which is causing the problem,
since __READ_ONCE_SIZE has casted that away and just uses "void *", but you
get the idea.

Will