Messages in this thread Patch in this message | | | From | Frederic Weisbecker <> | Subject | [PATCH] rcu/doc: Add a quick quiz to explain further why we need smp_mb__after_unlock_lock() | Date | Thu, 10 Jun 2021 17:50:29 +0200 |
| |
Add some missing critical pieces of explanation to understand the need for full memory barriers throughout the whole grace period state machine, thanks to Paul's explanations.
Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Uladzislau Rezki <urezki@gmail.com> Cc: Boqun Feng <boqun.feng@gmail.com> --- .../Tree-RCU-Memory-Ordering.rst | 33 +++++++++++++++++++ 1 file changed, 33 insertions(+)
diff --git a/Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst b/Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst index 11cdab037bff..f21432115627 100644 --- a/Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst +++ b/Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst @@ -112,6 +112,39 @@ on PowerPC. The ``smp_mb__after_unlock_lock()`` invocations prevent this ``WARN_ON()`` from triggering. ++-----------------------------------------------------------------------+ +| **Quick Quiz**: | ++-----------------------------------------------------------------------+ +| But the whole chain of rnp locking is enough for the readers to see | +| all the pre-grace-period accesses from the updater and for the updater| +| to see all the accesses from the readers performed before the end of | +| the grace period. So why do we need to enforce full ordering at all | +| through smp_mb__after_unlock_lock()? | ++-----------------------------------------------------------------------+ +| **Answer**: | ++-----------------------------------------------------------------------+ +| Because we still need to take care of the lockless counterparts of | +| RCU. The first key example here is grace period polling. Using | +| poll_state_synchronize_rcu() or cond_synchronize_rcu(), an updater | +| can rely solely on lockess full ordering to benefit from the usual | +| TREE RCU ordering guarantees. | +| | +| The second example lays behind the fact that a grace period still | +| claims to imply full memory ordering. Therefore in the following | +| scenario: | +| | +| CPU 0 CPU 1 | +| ---- ---- | +| WRITE_ONCE(X, 1) WRITE_ONCE(Y, 1) | +| synchronize_rcu() smp_mb() | +| r0 = READ_ONCE(Y) r1 = READ_ONCE(X) | +| | +| It must be impossible to have r0 == 0 && r1 == 0 after both CPUs | +| have completed their sequences, even if CPU 1 is in an RCU extended | +| quiescent state (idle mode) and thus won't report a quiescent state | +| throughout the common rnp locking chain. | ++-----------------------------------------------------------------------+ + This approach must be extended to include idle CPUs, which need RCU's grace-period memory ordering guarantee to extend to any RCU read-side critical sections preceding and following the current -- 2.25.1
| |