On Thu, Jun 10, 2004 at 13:23:50 -0400, Ed L Cashin wrote:
> Can anybody point out a specific problem with the (2.4) code below?
> When I saw it, it made me uneasy because I'd heard somewhat vague
> warnings against doing stuff like this. I can't see how the x86
> implementation could have a problem, but maybe some other architecture
> would?
Sure I can -- I have looked at actual implementations when I was
answering previous question....
> static struct mylock {
> spinlock_t lock;
> unsigned long flags;
> } mylock;
>
> void lock(struct mylock *p)
> {
> spin_lock_irqsave(&p->lock, p->flags);
> }
which is (because the #define of spin_lock_irqsave expands to it)
void lock(struct mylock *p) {
local_irq_save(p->flags);
spin_lock(&p->lock);
}
> void unlock(struct mylock *p)
> {
> spin_lock_irqrestore(&p->lock, p->flags);
> }
and this in turn is:
void unlock(struct mylock *p) {
spin_unlock(&p->lock);
local_irq_restore(&p->lock);
}
> void demonstration(void)
> {
> lock(&mylock);
> /* do stuff that doesn't sleep */
> unlock(&mylock);
> }
Now things start to happening...
CPU1 CPU2
(interrups enabled here) (interrupts disabled here)
local_irq_save(p->flags)
=> saved interrupts enabled
local_irq_save(p->flags)
=> saved interrupts disabled
spin_lock(&p->lock)
=> locked
spin_lock(&p->lock)
...
stuff is done... ...
spin_unlock(&p->lock) ...
=> unlocked ... still spinning
=> locked
local_irq_restore(p->flags)
=> interrupt STILL DISABLED!
stuff is done...
spin_unlock(&p->lock)
=> unlocked
local_irq_restore(p->flags)
=> interrupt disabled
CPU1 will never enable the interrupts...
I am not sure whether the race will happen on UP, but I think it will,
since (IIRC) interrupt handlers run with interrupts DISABLED. So, taking
into account that local_irq_save is pushfl; popl (flags); cli:
On the other hand, if you know the interrupt handler will run with
disabled interrupts, you won't irqsave/irqrestore (but if the function
is sometimes called from interrupt and sometimes not, you will ... and
lock up)
pushfl
popl (p->flags)
=> so now p->flags contain enabled interrupts...
<-- INTERRUPT REQUEST COMES IN
interrupts are now disabled
pushfl
popl (p->flags)
=> but now p->flags contain DISABLED interrupts...
cli
=> nothing happens...
lock is locked OK, sutuff gets done, lock is released
pushl (p->flags)
popfl
=> interrupts still disabled...
--> INTERRUPT REQUEST FINISHED
interrupts were enabled, iret
cli
=> interrupts disabled
lock is locked OK, stuff gets done, lock is released
pushl (p->flags)
=> still contains the DISABLED interrupts...
popfl
=> interrupts still disabled...
CPU will never enable interrupts....
-------------------------------------------------------------------------------
Jan 'Bulb' Hudec <bulb@ucw.cz>
Attachment:
signature.asc
Description: Digital signature
