On Mon, Aug 11, 2008 at 01:04:00PM +0200, Peter Zijlstra wrote:
One way to break the deadlock within the RCU subsystem would be something
similar to the following untested (known not to compile) patch. The
idea is that RCU detects that call_rcu() is being called from printk(), and
simply enqueues the callback in this case. For this to really work, RCU
needs something exported from printk() to allow it to make this
decision. I chose the static variable printk_cpu below just to present
the general idea.
Thoughts?
Buggy, but otherwise:
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
---
rcupreempt.c | 11 ++++++++---
1 file changed, 8 insertions(+), 3 deletions(-)
diff -urpNa -X dontdiff linux-2.6.27-rc1/kernel/rcupreempt.c linux-2.6.27-rc1-printk/kernel/rcupreempt.c
--- linux-2.6.27-rc1/kernel/rcupreempt.c 2008-07-30 08:48:17.000000000 -0700
+++ linux-2.6.27-rc1-printk/kernel/rcupreempt.c 2008-08-11 09:02:03.000000000 -0700
@@ -1118,17 +1118,22 @@ void call_rcu(struct rcu_head *head, voi
{
unsigned long flags;
struct rcu_data *rdp;
+ int this_cpu = smp_processor_id();
head->func = func;
head->next = NULL;
local_irq_save(flags);
rdp = RCU_DATA_ME();
- spin_lock(&rdp->lock);
- __rcu_advance_callbacks(rdp);
+ if (this_cpu == printk_cpu) {
+ spin_lock(&rdp->lock);
+ __rcu_advance_callbacks(rdp);
+ }
*rdp->nexttail = head;
rdp->nexttail = &head->next;
RCU_TRACE_RDP(rcupreempt_trace_next_add, rdp);
- spin_unlock_irqrestore(&rdp->lock, flags);
+ if (this_cpu == printk_cpu) {
+ spin_unlock_irqrestore(&rdp->lock, flags);
+ }
}
EXPORT_SYMBOL_GPL(call_rcu);
--
