On Wed, May 23, 2012 at 10:44 PM, loody [off-list ref] wrote:

take multi-core system for example, it is possible your completion
handler execute at the same time, except you add a spin lock for
activating your completion function.

That's the reason I'm asking in first place.
But if that is true, then it seems to me some drivers wouldn't be
working (unlikely, right?)

You can take a look at pwc completion for instance:

static void pwc_isoc_handler(struct urb *urb)
{
[snip]
        if (pdev->fill_buf == NULL)
                pdev->fill_buf = pwc_get_next_fill_buf(pdev);
---

As you can see, there is no lock here (though there is lock
to get the next fill_buf) and pdev->fill_buf is obviously not local.

On the other hand, em28xx has a tighter locking scheme:

static void em28xx_irq_callback(struct urb *urb)
{
[snip]
        /* Copy data from URB */
        spin_lock(&dev->slock);
        dev->isoc_ctl.isoc_copy(dev, urb);
        spin_unlock(&dev->slock);
---

Currently, I'm running a three core (yes three!) streaming at a high
rate and I'm following pwc locking (per queue). So far no races found (phew).

As a sidenote, em28xx locking can be improved for a very simple
reason: it's locking code instead of data.
Unfortunately, I don't own that hardware to fix and test such a patch.