At Mon, 22 Jun 2015 14:57:29 +0100, Mark Brown wrote:
On Mon, Jun 22, 2015 at 01:24:02PM +0200, Takashi Iwai wrote:
At Mon, 22 Jun 2015 11:30:34 +0100, Mark Brown wrote:
On Mon, Jun 22, 2015 at 11:58:24AM +0200, Takashi Iwai wrote:
And, looking at the code, it seems calling runtime suspend in the following way at probe:
pm_runtime_enable(&pdev->dev); if (!pm_runtime_enabled(&pdev->dev)) { ret = mtk_afe_runtime_resume(&pdev->dev); if (ret) goto err_pm_disable; }I'm confused, where's the call to runtime suspend?
It's in static const struct dev_pm_ops mtk_afe_pm_ops = { SET_RUNTIME_PM_OPS(mtk_afe_runtime_suspend, mtk_afe_runtime_resume, NULL) };
Sorry, I'm still confused about what you're seeing in the probe - I know where the callbacks for runtime PM are registered but I'm not seeing a call to suspend (or something that I'd expect to trigger one) in the above?
There is no place calling runtime suspend manually, that's why the compiler catches and warns.
But my concern above isn't about the warning itself. I just stumbled on the code invoking runtime resume while looking at this warning, and wondered the behavior with CONFIG_PM=n.
Usually this kind of warning could be simply fixed by adding a proper ifdef. But, this driver calls runtime resume in the probe manually.
Sure, that's a fairly common pattern though?
Depends. The more common pattern seems to call pm_runtime_resume(). And this will skip the call of runtime PM when CONFIG_PM=n.
I'm not sure whether this really behaves correctly, especially when a kernel is built without CONFIG_PM.
Could you be more specific about the problem you're seeing? If runtime PM is disabled pm_runtime_enabled() will return false and we'll run through the resume path during probe() instead, otherwise we'll runtime resume whenever we need to use the hardware.
The runtime suspend is never called when CONFIG_PM=n. OTOH, we call runtime resume *always* at probe when CONFIG_PM=n. This looks inconsistent to me.
Yeah, it should really resuspend the hardware in the remove path.
Right, then it'll be balanced properly.
If it's a part of the mandatory initialization, it should be named explicitly so, and make the runtime resume callback just calls it instead.
I disagree, I think either way is fine - if the clear intent and expectation is that the driver is used with runtime PM it seems fine to structure things to to say say "this is the special case path for !PM". I'd actually like to see this pattern better supported by the core so that drivers can enable runtime PM with calls that have !PM paths like in this driver, it'd make the whole !PM case a lot simpler.
I don't mind much in either way. But the current code looks somehow inconsistent.
Takashi