On 1/19/2015 11:44 AM, Russell King - ARM Linux wrote:
> To see why atomic_t is pure obfuscation:
>
> typedef struct {
> int counter;
> } atomic_t;
>
> So, counter is a plain int.
>
> On Mon, Jan 19, 2015 at 11:23:47AM -0800, Ray Jui wrote:
>> +static irqreturn_t bcm_iproc_i2c_isr(int irq, void *data)
>> +{
>> + struct bcm_iproc_i2c_dev *iproc_i2c = data;
>> + u32 status = readl(iproc_i2c->base + IS_OFFSET);
>> +
>> + status &= ISR_MASK;
>> +
>> + if (!status)
>> + return IRQ_NONE;
>> +
>> + writel(status, iproc_i2c->base + IS_OFFSET);
>> + atomic_set(&iproc_i2c->xfer_is_done, 1);
>
> #define atomic_set(v,i) (((v)->counter) = (i))
>
> So, this is the same as doing:
>
> iproc_i2c->xfer_is_done.counter = 1;
>
> which is merely setting the 'int' to 1.
>
>> + time_left = wait_for_completion_timeout(&iproc_i2c->done, time_left);
>> +
>> + /* disable all interrupts */
>> + writel(0, iproc_i2c->base + IE_OFFSET);
>> +
>> + if (!time_left && !atomic_read(&iproc_i2c->xfer_is_done)) {
>
> #define atomic_read(v) ACCESS_ONCE((v)->counter)
>
> This is practically the same as:
>
> if (!time_left && !iproc_i2c->xfer_is_done.counter) {
>
> except that this access will be guaranteed to happen just once at this
> location (see ACCESS_ONCE() in include/linux/compiler.h).
>
> However, complete()..wait_for_completion() ensures that there are
> barriers in the way: complete takes a spinlock on the waiter, so the
> write to iproc_i2c->xfer_is_done.counter will be visible by the time
> wait_for_completion() returns, and wait_for_completion() also does.
> The same spinlock is also manipulated by wait_for_completion(), which
> means there's barriers there as well, so it can't cache the value of
> "counter" across that call.
>
> So, the "volatile" access guaranteed by ACCESS_ONCE() isn't even
> needed here.
>
> (It would be needed if you were spinning in a loop, calling no other
> functions - but then you're supposed to use cpu_relax() in that
> circumstance, which has a compiler barrier in it, which ensures that
> it will re-read such a variable each time.)
>
I really learned a good lesson here. Thanks for the thorough explanation!
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
