On Friday, November 16, 2012 08:54:56 AM Huang Ying wrote:

On Fri, 2012-11-16 at 01:55 +0100, Rafael J. Wysocki wrote:

quoted

On Friday, November 16, 2012 01:44:00 AM Rafael J. Wysocki wrote:

quoted

On Friday, November 16, 2012 08:36:14 AM Huang Ying wrote:

quoted

On Thu, 2012-11-15 at 10:51 +0100, Rafael J. Wysocki wrote:

[...]

quoted

quoted

For this situation, if user "echo auto > .../power/control" for the
device, the runtime PM callbacks of device will be called.  I think that
is not intended.  So I think it is better to use some kind of flag or
state for that.

I'm not sure what situation exactly you have in mind.  Care to give an
exact scenario?

Ah, I see.  When we've just called drv->remove(), there is a window in
which user space may cause the driver's runtime PM callbacks to be
executed by changing its attribute to "auto".

So perhaps we should check pci_dev->driver rather than pci_dev->dev.driver
in the runtime PM callbacks?  With a few more changes that should allow us
to close that race.

Yes.  And I think, with pci_dev->driver (after some changes suggested by
Alan), we need not to use pm_runtime_get/put_skip_callbacks().

Good.  Can you please prepare a patch, then? :-)

Rafael


-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.