Re: [PATCH] driver core: flush async calls before testing driver removal

[Greg Kroah-Hartman]

On Sat, Dec 10, 2016 at 02:38:41PM +0200, Vladimir Zapolskiy wrote:
> Hello Greg,
> 
> On 12/10/2016 09:32 AM, Greg Kroah-Hartman wrote:
> > On Sat, Dec 10, 2016 at 02:15:19AM +0200, Vladimir Zapolskiy wrote:
> >> If CONFIG_DEBUG_TEST_DRIVER_REMOVE option is enabled a number of false
> >> positives are reported for ATA controller drivers, because ATA port
> >> probes are done asynchronously, and the same problem may also touch
> >> other asynchronously probed drivers.
> >>
> >> To reduce the rate of false reports on boot call async_synchronize_full()
> >> before attempting to remove a driver, the same is done in delete_module()
> >> syscall for all possible drivers and in __device_release_driver() function
> >> for asynchronously probed drivers.
> > 
> > __device_release_driver() already calls this function, why call it
> > again?
> > 
> 
> __device_release_driver() is not called on test removal of drivers, if
> CONFIG_DEBUG_TEST_DRIVER_REMOVE option is enabled.
> 
> This opens a possibility to races like one I've discovered:
> 
>   https://www.spinics.net/lists/linux-ide/msg53473.html
> 
> Next, async_synchronize_full() from __device_release_driver() is not
> called in case of removal of ATA controller drivers, because
> driver_allows_async_probing(drv) return value is false.

Hm, how does this not also get hit if you unbind/bind/unbind/bind/etc.
from userspace as well?  I don't think this is a
CONFIG_DEBUG_TEST_DRIVER_REMOVE issue, but just that this option finds
the problem corner cases as you are finding out :)

thanks,

greg k-h