Re: [PATCH] 2.4.18 raid1 - fix SMP locking/interrupt

Re: [PATCH] 2.4.18 raid1 - fix SMP locking/interrupt

[Manfred Spraul]

>
> However a bare spin_unlock_irq() in a function means that
> callers which wish to keep interrupts disabled are subtly
> subverted.   We've had bugs from this before.
>
It is trivial to catch such bugs at runtime. I tried it a year ago, and
immediately run into sleep_on() users that legitimately call
spin_lock_irq() with disabled interrupts. Perhaps they are gone now,
I'll retest my patch.

> So the irqrestore functions are much more robust.  I believe
> that they should be the default choice.  The non-restore
> versions should be viewed as a micro-optimised version,
> to be used with caution.  The additional expense of the save/restore
> is quite tiny - 20-30 cycles, perhaps.

OTHO, if a function doesn't work correctly if it's called with disabled
interrupts, then it should not use spin_lock_irqsave() - it's
misleading.
e.g. if it calls kmalloc(GFP_KERNEL), down(), schedule(), etc.

--
    Manfred

Re: [PATCH] 2.4.18 raid1 - fix SMP locking/interrupt

[Andrew Morton]

Manfred Spraul wrote:
> 
> >
> > However a bare spin_unlock_irq() in a function means that
> > callers which wish to keep interrupts disabled are subtly
> > subverted.   We've had bugs from this before.
> >
> It is trivial to catch such bugs at runtime. I tried it a year ago, and
> immediately run into sleep_on() users that legitimately call
> spin_lock_irq() with disabled interrupts. Perhaps they are gone now,
> I'll retest my patch.

heh.  Yes, I tried that too a year or so ago.  Basically just
add

	if (!(eflags & 0x0200))
		whine();

to spin_lock_irq().  There was a *lot* of whining.

> > So the irqrestore functions are much more robust.  I believe
> > that they should be the default choice.  The non-restore
> > versions should be viewed as a micro-optimised version,
> > to be used with caution.  The additional expense of the save/restore
> > is quite tiny - 20-30 cycles, perhaps.
> 
> OTHO, if a function doesn't work correctly if it's called with disabled
> interrupts, then it should not use spin_lock_irqsave() - it's
> misleading.
> e.g. if it calls kmalloc(GFP_KERNEL), down(), schedule(), etc.

mm?  Those are legal (albeit unpleasant) inside local_irq_save(),
but illegal inside global_cli() in 2.5.  Aren't they?  If not,
then release_kernel_lock() needs talking to.

-

Re: [PATCH] 2.4.18 raid1 - fix SMP locking/interrupt

[Manfred Spraul]

From: "Andrew Morton" <akpm@zip.com.au>
> > OTHO, if a function doesn't work correctly if it's called with
disabled
> > interrupts, then it should not use spin_lock_irqsave() - it's
> > misleading.
> > e.g. if it calls kmalloc(GFP_KERNEL), down(), schedule(), etc.
>
> mm?  Those are legal (albeit unpleasant) inside local_irq_save(),
> but illegal inside global_cli() in 2.5.  Aren't they?  If not,
> then release_kernel_lock() needs talking to.
>
If a function is called with disabled interrupts, then the caller
probably expects that the interrupts remain disabled - otherwise he
would have reenabled them before calling. schedule reenables interrupts.
The calls might be legal, but usually they indicate a bug.

--
    Manfred