Using random in interrupts for RT

Using random in interrupts for RT

[Steven Rostedt]

Thomas,

Ben Hutchings asked me if we still need "genirq: Disable random call on
preempt-rt" for -rt? With commit 902c098a366 "random: use lockless
techniques in the interrupt path" there is no more locks used. But does
it still produce high latencies?

-- Steve

Re: Using random in interrupts for RT

[Thomas Gleixner]

On Tue, 14 Aug 2012, Steven Rostedt wrote:
> Ben Hutchings asked me if we still need "genirq: Disable random call on
> preempt-rt" for -rt? With commit 902c098a366 "random: use lockless
> techniques in the interrupt path" there is no more locks used. But does
> it still produce high latencies?

Haven't tested it. Will put it on my ever growing todo list....

Re: Using random in interrupts for RT

[Steven Rostedt]

On Tue, 2012-08-14 at 10:30 -0400, Steven Rostedt wrote:
> Thomas,
> 
> Ben Hutchings asked me if we still need "genirq: Disable random call on
> preempt-rt" for -rt? With commit 902c098a366 "random: use lockless
> techniques in the interrupt path" there is no more locks used. But does
> it still produce high latencies?

Ben,

Looks that the #ifndef can't be removed just yet. The code is not
totally lockless...


add_interrupt_randomness ->
	credit_entropy_bits ->
		kill_fasync ->
			kill_fasync_rcu ->
				spin_lock_irqsave(&fa->fa_lock, flags)

to make things worse, that spinlock turns into a mutex on -rt, so it may
crash the box if triggered.

-- Steve

Re: Using random in interrupts for RT

[Ben Hutchings]

On Tue, Aug 21, 2012 at 02:12:13PM -0400, Steven Rostedt wrote:
> On Tue, 2012-08-14 at 10:30 -0400, Steven Rostedt wrote:
> > Thomas,
> > 
> > Ben Hutchings asked me if we still need "genirq: Disable random call on
> > preempt-rt" for -rt? With commit 902c098a366 "random: use lockless
> > techniques in the interrupt path" there is no more locks used. But does
> > it still produce high latencies?
> 
> Ben,
> 
> Looks that the #ifndef can't be removed just yet. The code is not
> totally lockless...
> 
> 
> add_interrupt_randomness ->
> 	credit_entropy_bits ->
> 		kill_fasync ->
> 			kill_fasync_rcu ->
> 				spin_lock_irqsave(&fa->fa_lock, flags)
> 
> to make things worse, that spinlock turns into a mutex on -rt, so it may
> crash the box if triggered.
 
Thanks for checking this.

Ben.

-- 
Ben Hutchings
We get into the habit of living before acquiring the habit of thinking.
                                                              - Albert Camus