Dropping Feng Kan <f...@apm.com> and Toan Le <toa...@apm.com> because
their mails are bouncing.
On Thu, Jun 25, 2020 at 8:19 AM Rafael J. Wysocki <raf...@kernel.org> wrote:
>
> On Thu, Jun 25, 2020 at 5:24 AM Saravana Kannan <sarava...@google.com> wrote:
> >
> > Under the following conditions:
> > - driver A is built in and can probe device-A
> > - driver B is a module and can probe device-B
> > - device-A is supplier of device-B
> >
> > Without this patch:
> > 1. device-A is added.
> > 2. device-B is added.
> > 3. dpm_list is now [device-A, device-B].
> > 4. driver-A defers probe of device-A.
> > 5. deferred probe of device-A is reattempted
> > 6. device-A is moved to end of dpm_list.
> > 6. dpm_list is now [device-B, device-A].
> > 7. driver-B is loaded and probes device-B.
> > 8. dpm_list stays as [device-B, device-A].
> >
> > Suspend (which goes in the reverse order of dpm_list) fails because
> > device-A (supplier) is suspended before device-B (consumer).
> >
> > With this patch:
> > 1. device-A is added.
> > 2. device-B is added.
> > 3. dpm_list is now [device-A, device-B].
> > 4. driver-A defers probe of device-A.
> > 5. deferred probe of device-A is reattempted later.
> > 6. dpm_list is now [device-B, device-A].
> > 7. driver-B is loaded and probes device-B.
> > 8. dpm_list is now [device-A, device-B].
> >
> > Suspend works because device-B (consumer) is suspended before device-A
> > (supplier).
> >
> > Fixes: 494fd7b7ad10 ("PM / core: fix deferred probe breaking suspend resume
> > order")
> > Fixes: 716a7a259690 ("driver core: fw_devlink: Add support for batching
> > fwnode parsing")
> > Cc: Geert Uytterhoeven <ge...@linux-m68k.org>
> > Signed-off-by: Saravana Kannan <sarava...@google.com>
> > ---
> > drivers/base/dd.c | 16 ++++++++++++++++
> > 1 file changed, 16 insertions(+)
> >
> > diff --git a/drivers/base/dd.c b/drivers/base/dd.c
> > index 9a1d940342ac..52b2148c7983 100644
> > --- a/drivers/base/dd.c
> > +++ b/drivers/base/dd.c
> > @@ -109,6 +109,8 @@ static void deferred_probe_work_func(struct work_struct
> > *work)
> > * probe makes that very unsafe.
> > */
> > device_pm_move_to_tail(dev);
> > + /* Greg/Rafael: SHOULD I DELETE THIS? ^^ I think I should,
> > but
> > + * I'm worried if it'll have some unintended consequeneces.
> > */
>
> Yes, this needs to go away if you make the other change.
>
> >
> > dev_dbg(dev, "Retrying from deferred list\n");
> > bus_probe_device(dev);
> > @@ -557,6 +559,20 @@ static int really_probe(struct device *dev, struct
> > device_driver *drv)
> > goto re_probe;
> > }
> >
> > + /*
> > + * The devices are added to the dpm_list (resume/suspend (reverse
> > + * order) list) as they are registered with the driver core. But the
> > + * order the devices are added doesn't necessarily match the real
> > + * dependency order.
> > + *
> > + * The successful probe order is a much better signal. If a device
> > just
> > + * probed successfully, then we know for sure that all the devices
> > that
> > + * probed before it don't depend on the device. So, we can safely
> > move
> > + * the device to the end of the dpm_list. As more devices probe,
> > + * they'll automatically get ordered correctly.
> > + */
> > + device_pm_move_to_tail(dev);
>
> But it would be good to somehow limit this to the devices affected by
> deferred probing or we'll end up reordering dpm_list unnecessarily for
> many times in the actual majority of cases.
Yes, lots of unnecessary reordering, but doing it only for deferred
probes IS the problem. In the example I gave, the consumer is never
deferred probe because the supplier happens to finish probing before
the consumer probe is even attempted.
I'm open to other suggestions, but I think this is needed for all the
cases or at least more cases to be handled correctly.
One alternative I was thinking was not adding the device to the
dpm_list until it's probed. But I have the vague feeling of other
things between device_add() and device probe that expect the device to
be in the dpm_list.
-Saravana
