• Linus Torvalds's avatar
    rcu: locking and unlocking need to always be at least barriers · f0d1e74c
    Linus Torvalds authored
    commit 66be4e66a7f422128748e3c3ef6ee72b20a6197b upstream.
    
    Herbert Xu pointed out that commit bb73c52b ("rcu: Don't disable
    preemption for Tiny and Tree RCU readers") was incorrect in making the
    preempt_disable/enable() be conditional on CONFIG_PREEMPT_COUNT.
    
    If CONFIG_PREEMPT_COUNT isn't enabled, the preemption enable/disable is
    a no-op, but still is a compiler barrier.
    
    And RCU locking still _needs_ that compiler barrier.
    
    It is simply fundamentally not true that RCU locking would be a complete
    no-op: we still need to guarantee (for example) that things that can
    trap and cause preemption cannot migrate into the RCU locked region.
    
    The way we do that is by making it a barrier.
    
    See for example commit 386afc91 ("spinlocks and preemption points
    need to be at least compiler barriers") from back in 2013 that had
    similar issues with spinlocks that become no-ops on UP: they must still
    constrain the compiler from moving other operations into the critical
    region.
    
    Now, it is true that a lot of RCU operations already use READ_ONCE() and
    WRITE_ONCE() (which in practice likely would never be re-ordered wrt
    anything remotely interesting), but it is also true that that is not
    globally the case, and that it's not even necessarily always possible
    (ie bitfields etc).
    Reported-by: default avatarHerbert Xu <herbert@gondor.apana.org.au>
    Fixes: bb73c52b ("rcu: Don't disable preemption for Tiny and Tree RCU readers")
    Cc: stable@kernel.org
    Cc: Boqun Feng <boqun.feng@gmail.com>
    Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
    Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
    Signed-off-by: default avatarGreg Kroah-Hartman <gregkh@linuxfoundation.org>
    f0d1e74c
rcupdate.h 40.2 KB