Re: [PATCH] rcutorture: Fix rcu_torture_pipe_update_one()/rcu_torture_writer() data race and concurrency bug
From: Mathieu Desnoyers
Date: Wed Mar 06 2024 - 22:06:31 EST
On 2024-03-06 21:43, Linus Torvalds wrote:
[...]
Honestly, this all makes me think that we'd be *much* better off
showing the real "handoff" with smp_store_release() and
smp_load_acquire().
We've done something similar in liburcu (userspace code) to allow
Thread Sanitizer to understand the happens-before relationships
within the RCU implementations and lock-free data structures.
Moving to load-acquire/store-release (C11 model in our case)
allowed us to provide enough happens-before relationship for
Thread Sanitizer to understand what is happening under the
hood in liburcu and perform relevant race detection of user
code.
As far as the WRITE_ONCE(x, READ_ONCE(x) + 1) pattern
is concerned, the only valid use-case I can think of is
split counters or RCU implementations where there is a
single updater doing the increment, and one or more
concurrent reader threads that need to snapshot a
consistent value with READ_ONCE().
Thanks,
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com