On Thu, Dec 14, 2023 at 03:09:36PM +0100, Johan Hovold wrote:
> On Thu, Dec 14, 2023 at 04:04:49PM +0200, Dmitry Baryshkov wrote:
> > On Thu, 14 Dec 2023 at 16:01, Johan Hovold <johan@xxxxxxxxxx> wrote:
> > > On Thu, Dec 14, 2023 at 01:04:43PM +0200, Dmitry Baryshkov wrote:
> > > > This is a known problem. Successful probes during the probe deferral
> > > > loop causes the whole loop to be reiterated. Creating child devices
> > > > usually results in a successful probe. Aso I thought that just
> > > > creating new device also causes a reprobe, but I can not find any
> > > > evidence now.
> > >
> > > This still needs to be described in the commit message.
> > >
> > > Only a successful probe should trigger a reprobe, and when the child
> > > devices are registered the parent is not yet on the deferred probe list.
> > > So something is not right or missing here.
> >
> > Child devices can be successfully probed, then the parent gets
> > -EPROBE_DEFER, removes children and then it goes on and on.
>
> So what? As I described above, the successful probe of the children
> should have nothing to do with whether the parent is reprobed.
>
> If that isn't the case, then explain how.
I took a closer look at this and indeed we do have code that triggers a
reprobe of a device in case there was a successful probe while the
device was probing.
This was introduced by commit 58b116bce136 ("drivercore: deferral race
condition fix") and the workaround for the reprobe-loop bug that hack
led to is to not return -EPROBE_DEFER after registering child devices as
no one managed to come up with a proper fix. This was documented here:
fbc35b45f9f6 ("Add documentation on meaning of -EPROBE_DEFER")
But please spell this out in some more detail in the commit message, and
add a Fixes and CC stable tag.
Johan
