Akinobu Mita (on Wed, 25 Jan 2006 20:32:06 +0900) wrote:
>o generic {,test_and_}{set,clear,change}_bit() (atomic bitops)
...
>+static __inline__ void set_bit(int nr, volatile unsigned long *addr)
>+{
>+ unsigned long mask = BITOP_MASK(nr);
>+ unsigned long *p = ((unsigned long *)addr) + BITOP_WORD(nr);
>+ unsigned long flags;
>+
>+ _atomic_spin_lock_irqsave(p, flags);
>+ *p |= mask;
>+ _atomic_spin_unlock_irqrestore(p, flags);
>+}
Be very, very careful about using these generic *_bit() routines if the
architecture supports non-maskable interrupts.
NMI events can occur at any time, including when interrupts have been
disabled by *_irqsave(). So you can get NMI events occurring while a
*_bit fucntion is holding a spin lock. If the NMI handler also wants
to do bit manipulation (and they do) then you can get a deadlock
between the original caller of *_bit() and the NMI handler.
Doing any work that requires spinlocks in an NMI handler is just asking
for deadlock problems. The generic *_bit() routines add a hidden
spinlock behind what was previously a safe operation. I would even say
that any arch that supports any type of NMI event _must_ define its own
bit routines that do not rely on your _atomic_spin_lock_irqsave() and
its hash of spinlocks.
