[PATCH] kernel/rcupdate.c microcleanup

[PATCH] kernel/rcupdate.c microcleanup

[Andrey Panin]

[-- Attachment #1: Type: text/plain, Size: 469 bytes --]

Hi all,

attached patch (2.5.64) contains small cleanup of RCU code:
    - move smp_processor_id() outside of irq disabled region in call_rcu();
    - consolidate multiple spin_unlock() in the rcu_check_quiescent_state(),
      remove some unneeded {} and make this function inline.

Tested and works (at least doesn't crash). Please consider applying.

Best regards.

-- 
Andrey Panin		| Embedded systems software developer
pazke@orbita1.ru	| PGP key: wwwkeys.pgp.net

[-- Attachment #2: patch-rcu --]
[-- Type: text/plain, Size: 1788 bytes --]

diff -urN -X /usr/share/dontdiff linux-2.5.64.vanilla/kernel/rcupdate.c linux-2.5.64/kernel/rcupdate.c
--- linux-2.5.64.vanilla/kernel/rcupdate.c	Thu Nov 28 01:35:46 2002
+++ linux-2.5.64/kernel/rcupdate.c	Mon Mar 10 20:18:48 2003
@@ -67,13 +67,12 @@
  */
 void call_rcu(struct rcu_head *head, void (*func)(void *arg), void *arg)
 {
-	int cpu;
+	int cpu = smp_processor_id();
 	unsigned long flags;
 
 	head->func = func;
 	head->arg = arg;
 	local_irq_save(flags);
-	cpu = smp_processor_id();
 	list_add_tail(&head->list, &RCU_nxtlist(cpu));
 	local_irq_restore(flags);
 }
@@ -117,13 +116,12 @@
  * switch). If so and if it already hasn't done so in this RCU
  * quiescent cycle, then indicate that it has done so.
  */
-static void rcu_check_quiescent_state(void)
+static inline void rcu_check_quiescent_state(void)
 {
 	int cpu = smp_processor_id();
 
-	if (!test_bit(cpu, &rcu_ctrlblk.rcu_cpu_mask)) {
+	if (!test_bit(cpu, &rcu_ctrlblk.rcu_cpu_mask))
 		return;
-	}
 
 	/* 
 	 * Races with local timer interrupt - in the worst case
@@ -134,23 +132,22 @@
 		RCU_last_qsctr(cpu) = RCU_qsctr(cpu);
 		return;
 	}
-	if (RCU_qsctr(cpu) == RCU_last_qsctr(cpu)) {
+	if (RCU_qsctr(cpu) == RCU_last_qsctr(cpu))
 		return;
-	}
 
 	spin_lock(&rcu_ctrlblk.mutex);
-	if (!test_bit(cpu, &rcu_ctrlblk.rcu_cpu_mask)) {
-		spin_unlock(&rcu_ctrlblk.mutex);
-		return;
-	}
+	if (!test_bit(cpu, &rcu_ctrlblk.rcu_cpu_mask))
+		goto out_unlock;
+
 	clear_bit(cpu, &rcu_ctrlblk.rcu_cpu_mask);
 	RCU_last_qsctr(cpu) = RCU_QSCTR_INVALID;
-	if (rcu_ctrlblk.rcu_cpu_mask != 0) {
-		spin_unlock(&rcu_ctrlblk.mutex);
-		return;
-	}
+	if (rcu_ctrlblk.rcu_cpu_mask != 0)
+		goto out_unlock;
+
 	rcu_ctrlblk.curbatch++;
 	rcu_start_batch(rcu_ctrlblk.maxbatch);
+
+out_unlock:
 	spin_unlock(&rcu_ctrlblk.mutex);
 }
 

Re: [PATCH] kernel/rcupdate.c microcleanup

[Brian Gerst]

Andrey Panin wrote:
> Hi all,
> 
> attached patch (2.5.64) contains small cleanup of RCU code:
>     - move smp_processor_id() outside of irq disabled region in call_rcu();
>     - consolidate multiple spin_unlock() in the rcu_check_quiescent_state(),
>       remove some unneeded {} and make this function inline.
> 
> Tested and works (at least doesn't crash). Please consider applying.
> 
> Best regards.
> 
> 
> 
> ------------------------------------------------------------------------
> 
> diff -urN -X /usr/share/dontdiff linux-2.5.64.vanilla/kernel/rcupdate.c linux-2.5.64/kernel/rcupdate.c
> --- linux-2.5.64.vanilla/kernel/rcupdate.c	Thu Nov 28 01:35:46 2002
> +++ linux-2.5.64/kernel/rcupdate.c	Mon Mar 10 20:18:48 2003
> @@ -67,13 +67,12 @@
>   */
>  void call_rcu(struct rcu_head *head, void (*func)(void *arg), void *arg)
>  {
> -	int cpu;
> +	int cpu = smp_processor_id();
>  	unsigned long flags;
>  
>  	head->func = func;
>  	head->arg = arg;
>  	local_irq_save(flags);
> -	cpu = smp_processor_id();
>  	list_add_tail(&head->list, &RCU_nxtlist(cpu));
>  	local_irq_restore(flags);
>  }

This is not preempt-safe.  smp_processor_id() can only be used in a 
preempt-disabled region (which local_irq_save() provides).

--
				Brian Gerst

Re: [PATCH] kernel/rcupdate.c microcleanup

[Linus Torvalds]

On Tue, 11 Mar 2003, Andrey Panin wrote:
> 
> attached patch (2.5.64) contains small cleanup of RCU code:
>     - move smp_processor_id() outside of irq disabled region in call_rcu();

That looks wrong, and is likely to break with preemption, since the caller 
could potentially be moved to another CPU between the smp_processor_id() 
and the irq disable.

		Linus