[question] Can the execution of the atomtic operation instruction pair lwarx/stwcx be interrrupted by local HW interruptions?

wyang w90p710 at gmail.com
Tue Jan 7 12:00:20 EST 2014

On 01/07/2014 06:05 AM, Scott Wood wrote:
> On Mon, 2014-01-06 at 13:27 +0800, wyang wrote:
>> On 01/06/2014 11:41 AM, Gavin Hu wrote:
>>> Thanks your response.  :)
>>> But that means that these optimitive operations like atomic_add()
>>> aren't optimitive actully in PPC architecture, right? Becuase they
>>> can be interrupted by loacl HW interrupts. Theoretically, the ISR
>>> also can access the atomic gloable variable.
>> Nope, my understand is that if you wanna sync kernel primitive code
>> with ISR, you have responsibility to disable local interrupts.
>> atomic_add does not guarantee to handle such case.
> atomic_add() and other atomics do handle that case.  Interrupts are not
> disabled, but there's a stwcx. in the interrupt return code to make sure
> the reservation gets cleared.

Yeah, Can you provide more detail info about why they can handle that 
case? The following is my understand:

Let us assume that there is a atomic global variable(var_a) and its 
initial value is 0.

The kernel attempts to execute atomic_add(1, var_a), after lwarx a async 
interrupt happens, and the ISR also accesses "var_a" variable and 
executes atomic_add.

static __inline__ void atomic_add(int a, atomic_t *v)
     int t;

     __asm__ __volatile__(
"1:    lwarx    %0,0,%3        # atomic_add\n\
----------------------------------  <----------- interrupt 
happens------->        ISR also operates this global variable "var_a" 
such as also executing atomic_add(1, var_a). so the
               var_a would is 1.
     add    %0,%2,%0\n"
"    stwcx.    %0,0,%3 \n\ <----- After interrupt code returns, the 
reservation is cleared. so CR0 is not equal to 0, and then jump the 1 
label. the var_a will be 2.
     bne-    1b"
     : "=&r" (t), "+m" (v->counter)
     : "r" (a), "r" (&v->counter)
     : "cc");

So the value of var_a is 2 rather than 1. Thats why i said that 
atomic_add does not handle such case. If I miss something, please 
correct me.:-)

> -Scott

More information about the Linuxppc-dev mailing list