Re: kmalloc without GFP_xxx?

[Denis Vlasenko <vda@ilport.com.ua>]

On Wednesday 29 June 2005 16:44, Steven Rostedt wrote:
> 
> On Wed, 29 Jun 2005, Arjan van de Ven wrote:
> > >
> > > but it sets irqs_disabled() IIRC.
> >
> > only spin_lock_irq() and co do.
> > not the simple spin_lock()
> >
> 
> It may be dangerous to use spin_lock with interrupts enabled, since you
> have to make sure that no interrupt ever grabs that lock.  Although I do
> recall seeing a few locks like this.  But even so, you can transfer the
> latency of the interrupts going off while holding that lock to another CPU
> which IMHO is a bad thing.  Also a simple spin_lock would disable

This is why I always use _irqsave. Less error prone.
And locking is a very easy to get 'slightly' wrong, thus
I trade 0.1% of performance for code simplicity.

> preemption with CONFIG_PREEMPT set and that would make in_atomic fail.
> But to implement a kmalloc_auto you would always need to have a preempt
> count.
> 
> I'm not for a kmalloc_auto, but something like it would be useful for a
> function that can work for either context, and just fail nicely if the
> ATOMIC is set and the malloc can't get memory.  A function like this would
> currently have to always use ATOMIC even if it could have used KERNEL for
> some scenarios, since it would suffer the same pitfalls as a kmalloc_auto
> in determining its context.

This is more or less what I meant. Why think about each kmalloc and when you
eventually did get it right: "Aha, we _sometimes_ get called from spinlocked code,
GFP_ATOMIC then" - you still do atomic alloc even if cases when you
were _not_ called from locked code! Thus you needed to think longer and got
code which is worse.
--
vda


-
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/