spinlock_irqsave() && flags (Was: pm80xx: Spinlock fix)

[Oleg Nesterov <oleg@redhat.com>]

On 12/23, Oleg Nesterov wrote:
>
> Perhaps we should ask the maintainers upstream? Even if this works, I am
> not sure this is _supposed_ to work. I mean, in theory spin_lock_irqave()
> can be changed as, say
>
> 	#define spin_lock_irqsave(lock, flags)		\
> 		do {    				\
> 			local_irq_save(flags);		\
> 			spin_lock(lock);		\
> 		} while (0)
>
> (and iirc it was defined this way a long ago). In this case "flags" is
> obviously not protected.

Yes, lets ask the maintainers.

In short, is this code

	spinlock_t LOCK;
	unsigned long FLAGS;

	void my_lock(void)
	{
		spin_lock_irqsave(&LOCK, FLAGS);
	}

	void my_unlock(void)
	{
		spin_unlock_irqrestore(&LOCK, FLAGS);
	}

correct or not?

Initially I thought that this is obviously wrong, irqsave/irqrestore
assume that "flags" is owned by the caller, not by the lock. And iirc
this was certainly wrong in the past.

But when I look at spinlock.c it seems that this code can actually work.
_irqsave() writes to FLAGS after it takes the lock, and _irqrestore()
has a copy of FLAGS before it drops this lock.

And it turns out, some users assume this should work, for example

	arch/arm/mach-omap2/powerdomain.c:
		pwrdm_lock() and pwrdm_unlock()

	drivers/net/wireless/brcm80211/brcmfmac/fwsignal.c:
		brcmf_fws_lock() and brcmf_fws_unlock()

seem to do exactly this. Plus the pending patch for drivers/scsi/pm8001/.

So is it documented somewhere that this sequence is correct, or the code
above should be changed even if it happens to work?

Oleg.