Re: [PATCH] local_irq_disable removal

["Mika Penttilä" <mika.penttila@kolumbus.fi>]

Sven-Thorsten Dietrich wrote:

>On Sat, 2005-06-11 at 18:00 +0300, Mika Penttilä wrote:
>  
>
>>Ingo Molnar wrote:
>>
>>    
>>
>>>i've done two more things in the latest patches:
>>>
>>>- decoupled the 'soft IRQ flag' from the hard IRQ flag. There's
>>> basically no need for the hard IRQ state to follow the soft IRQ state. 
>>> This makes the hard IRQ disable primitives a bit faster.
>>>
>>>- for raw spinlocks i've reintroduced raw_local_irq primitives again.
>>> This helped get rid of some grossness in sched.c, and the raw
>>> spinlocks disable preemption anyway. It's also safer to just assume
>>> that if a raw spinlock is used together with the IRQ flag that the
>>> real IRQ flag has to be disabled.
>>>
>>>these changes dont really impact scheduling/preemption behavior, they 
>>>are cleanup/robustization changes.
>>>
>>>	Ingo
>>>-
>>>To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
>>>the body of a message to majordomo@vger.kernel.org
>>>More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>>Please read the FAQ at  http://www.tux.org/lkml/
>>>
>>> 
>>>
>>>      
>>>
>>With the soft IRQ flag local_irq_disable() doesn't seem to protect 
>>against soft interrupts (via SA_NODELAY interrupt-> invoke_softirq()). 
>>Could this be a problem?
>>    
>>
>
>Only if you run SOFT IRQs as SA_NODELAY, which is going to KILL all your
>preemption gains with the first arriving network packet.
>
>And that is, if you don't get buried in "scheduling while atomic" printk
>messages first.
>
>SA_NODELAY is not generally allowed in PREEMPT_RT, except for code
>designed to take advantage of the IRQ void that has been created. 
>
>This code must follow a new set of rules, which people who design RT
>apps are really happy to accet, they have to accept worse compromises
>with the alternatives (subkernel or ANOTHER OS (ugh))
>
>Sven
>
>
>
>  
>
The timer irq is run as NODELAY, so soft irqs are run against 
local_irq_disable sections all the time.

--Mika