On Wed, Jul 21, 2021 at 01:21:19PM -0700, Paul E. McKenney wrote:
> Accesses to ->qsmask are normally protected by ->lock, but there is an
> exception in the diagnostic code in rcu_check_boost_fail(). This commit
> therefore applies data_race() to this access to avoid KCSAN complaining
> about the C-language writes protected by ->lock.
>
> Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> ---
> kernel/rcu/tree_stall.h | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/rcu/tree_stall.h b/kernel/rcu/tree_stall.h
> index 42847caa3909b..6dd6c9aa3f757 100644
> --- a/kernel/rcu/tree_stall.h
> +++ b/kernel/rcu/tree_stall.h
> @@ -766,7 +766,7 @@ bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
>
> rcu_for_each_leaf_node(rnp) {
> if (!cpup) {
> - if (READ_ONCE(rnp->qsmask)) {
> + if (data_race(READ_ONCE(rnp->qsmask))) {
If the write sides allow normal writes, i.e. allowing store tearing, the
READ_ONCE() here could read incomplete writes, which could be anything
basically? And we get the same result if we remove the READ_ONCE(),
don't we? Yes, I know, without the READ_ONCE(), compilers can do
anything to optimize on rnp->qsmask, but the end result is same or
similar to reading incomplete writes (which is also a result by compiler
optimization). So if we mark something as data_race(), **in theory**, it
makes no difference with or without the READ_ONCE(), so I think maybe we
can remove the READ_ONCE() here?
Regards,
Boqun
> return false;
> } else {
> if (READ_ONCE(rnp->gp_tasks))
> --
> 2.31.1.189.g2e36527f23
>
