Re: [Qemu-devel] [PATCH 7/9] rcu: prod call_rcu thread when calling synchronize_rcu

[Fam Zheng]

On Tue, 02/03 13:52, Paolo Bonzini wrote:
> call_rcu operates on the principle that either there is a steady stream of
> incoming RCU callbacks, or it is not worthwhile to wake up and process the
> few that are there.
> 
> This however makes it hard to assert in testcases that all RCU callbacks
> are processed.  To avoid this, make call_rcu also process callbacks if there
> is a steady stream of synchronize_rcu calls.
> 
> This avoids deadlocks in the upcoming test-rcu-list unit test, which waits
> for call_rcu to reclaim all nodes that it allocates.  Especially with very
> high load on the host, call_rcu decided to wait for a few more callbacks
> to pile up, but the test was done and was not going to produce more.
> 
> Signed-off-by: Paolo Bonzini <address@hidden>
> ---
>  util/rcu.c | 8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/util/rcu.c b/util/rcu.c
> index c9c3e6e..aa9f639 100644
> --- a/util/rcu.c
> +++ b/util/rcu.c
> @@ -48,6 +48,9 @@ unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
>  QemuEvent rcu_gp_event;
>  static QemuMutex rcu_gp_lock;
>  
> +static int rcu_call_count;
> +static QemuEvent rcu_call_ready_event;
> +
>  /*
>   * Check whether a quiescent state was crossed between the beginning of
>   * update_counter_and_wait and now.
> @@ -149,6 +152,9 @@ void synchronize_rcu(void)
>      }
>  
>      qemu_mutex_unlock(&rcu_gp_lock);
> +    if (atomic_read(&rcu_call_count)) {
> +        qemu_event_set(&rcu_call_ready_event);
> +    }
>  }
>  
>  
> @@ -159,8 +165,6 @@ void synchronize_rcu(void)
>   */
>  static struct rcu_head dummy;
>  static struct rcu_head *head = &dummy, **tail = &dummy.next;
> -static int rcu_call_count;
> -static QemuEvent rcu_call_ready_event;
>  
>  static void enqueue(struct rcu_head *node)
>  {
> -- 
> 1.8.3.1
> 
> 

Reviewed-by: Fam Zheng <address@hidden>