2014-12-12 11:50:33

by Bill Huang

[permalink] [raw]
Subject: [RFC 1/1] driver core: re-order dpm_list after a succussful probe

The dpm_list was added in the call "device_add" and when we do defer
probe we'll explicitly move the probed device to be the last in the
dpm_list, we should do the same for the normal probe since there are
cases that we won't have chance to do defer probe to change the PM order
as the below example.

If we would like the device driver A to be suspended earlier than the
device driver B, we won't have chance to do defer probe to fix the
suspend dependency since at the time device driver A is probed, device B
was up and running.

Device A was added
Device B was added
Driver for device B was binded
Driver for device A was binded

Signed-off-by: Bill Huang <[email protected]>
---

It seems to me this is a bug in the core driver, but I'm not sure how should
we fix it.

- Do we have better fix?
- This proposed fix or any other fix might introduces side effect that breaks
existing working suspend dependencies which happen to work based on the
existing wrong suspend order.

Any thoughts? Thanks.

drivers/base/dd.c | 4 ++++
1 file changed, 4 insertions(+)

diff --git a/drivers/base/dd.c b/drivers/base/dd.c
index cdc779c..54886d2 100644
--- a/drivers/base/dd.c
+++ b/drivers/base/dd.c
@@ -308,6 +308,10 @@ static int really_probe(struct device *dev, struct device_driver *drv)
goto probe_failed;
}

+ device_pm_lock();
+ device_pm_move_last(dev);
+ device_pm_unlock();
+
driver_bound(dev);
ret = 1;
pr_debug("bus: '%s': %s: bound device %s to driver %s\n",
--
1.9.1


2014-12-12 15:34:06

by Greg Kroah-Hartman

[permalink] [raw]
Subject: Re: [RFC 1/1] driver core: re-order dpm_list after a succussful probe

On Fri, Dec 12, 2014 at 03:50:15AM -0800, Bill Huang wrote:
> The dpm_list was added in the call "device_add" and when we do defer
> probe we'll explicitly move the probed device to be the last in the
> dpm_list, we should do the same for the normal probe since there are
> cases that we won't have chance to do defer probe to change the PM order
> as the below example.
>
> If we would like the device driver A to be suspended earlier than the
> device driver B, we won't have chance to do defer probe to fix the
> suspend dependency since at the time device driver A is probed, device B
> was up and running.
>
> Device A was added
> Device B was added
> Driver for device B was binded
> Driver for device A was binded
>
> Signed-off-by: Bill Huang <[email protected]>
> ---
>
> It seems to me this is a bug in the core driver, but I'm not sure how should
> we fix it.
>
> - Do we have better fix?
> - This proposed fix or any other fix might introduces side effect that breaks
> existing working suspend dependencies which happen to work based on the
> existing wrong suspend order.
>
> Any thoughts? Thanks.
>
> drivers/base/dd.c | 4 ++++
> 1 file changed, 4 insertions(+)
>
> diff --git a/drivers/base/dd.c b/drivers/base/dd.c
> index cdc779c..54886d2 100644
> --- a/drivers/base/dd.c
> +++ b/drivers/base/dd.c
> @@ -308,6 +308,10 @@ static int really_probe(struct device *dev, struct device_driver *drv)
> goto probe_failed;
> }
>
> + device_pm_lock();
> + device_pm_move_last(dev);
> + device_pm_unlock();
> +
> driver_bound(dev);
> ret = 1;
> pr_debug("bus: '%s': %s: bound device %s to driver %s\n",


Adding Grant, as he did the deferred probe stuff...

And it's the middle of the merge window, I'll not have time to look at
this for a few weeks at the earliest, sorry.

thanks,

greg k-h

2014-12-18 08:06:16

by Bill Huang

[permalink] [raw]
Subject: Re: [RFC 1/1] driver core: re-order dpm_list after a succussful probe



On 12/17/2014 10:47 PM, Bibek Basu wrote:
> Hi Bill,
>
> Though I like your solution, I have a usecase where the driver probe
> sequence itself is not right. Both the driver are module_init but one
> depends on another during suspend sequence.
> In such a situation, my system hangs. What do you suggest to do in that
> case? Should I get my driver registration sequence right and how?
> Moving tegra-pcie driver above in the probe sequence by making the
> driver subsystem_initcall solved the issue I am facing with this patch.
> But I don't think that's allowed solution?

To change the probe sequence, use defer probe is the right choice.
>
> Example:
>
> Probe sequence:
> driver pcieport
> driver tegra-pcie
>
> Due to your patch, suspend_noirq for tegra_pcie will be called before

Are you sure? My change will only affect pm devices in dpm_list,
suspend_noirq should still be called after all devices in dpm_list were
suspended.

> pcieport. While pcieport tries to read through pci_bus_read_config_dword
> with clocks and power off to the pcie controller and eventually leads to
> a crash.
>
>
>
> regards
> Bibek
>
> On Fri, Dec 12, 2014 at 9:04 PM, Greg KH <[email protected]
> <mailto:[email protected]>> wrote:
>
> On Fri, Dec 12, 2014 at 03:50:15AM -0800, Bill Huang wrote:
> > The dpm_list was added in the call "device_add" and when we do defer
> > probe we'll explicitly move the probed device to be the last in the
> > dpm_list, we should do the same for the normal probe since there are
> > cases that we won't have chance to do defer probe to change the
> PM order
> > as the below example.
> >
> > If we would like the device driver A to be suspended earlier than the
> > device driver B, we won't have chance to do defer probe to fix the
> > suspend dependency since at the time device driver A is probed,
> device B
> > was up and running.
> >
> > Device A was added
> > Device B was added
> > Driver for device B was binded
> > Driver for device A was binded
> >
> > Signed-off-by: Bill Huang <[email protected]
> <mailto:[email protected]>>
> > ---
> >
> > It seems to me this is a bug in the core driver, but I'm not sure
> how should
> > we fix it.
> >
> > - Do we have better fix?
> > - This proposed fix or any other fix might introduces side effect
> that breaks
> > existing working suspend dependencies which happen to work
> based on the
> > existing wrong suspend order.
> >
> > Any thoughts? Thanks.
> >
> > drivers/base/dd.c | 4 ++++
> > 1 file changed, 4 insertions(+)
> >
> > diff --git a/drivers/base/dd.c b/drivers/base/dd.c
> > index cdc779c..54886d2 100644
> > --- a/drivers/base/dd.c
> > +++ b/drivers/base/dd.c
> > @@ -308,6 +308,10 @@ static int really_probe(struct device *dev,
> struct device_driver *drv)
> > goto probe_failed;
> > }
> >
> > + device_pm_lock();
> > + device_pm_move_last(dev);
> > + device_pm_unlock();
> > +
> > driver_bound(dev);
> > ret = 1;
> > pr_debug("bus: '%s': %s: bound device %s to driver %s\n",
>
>
> Adding Grant, as he did the deferred probe stuff...
>
> And it's the middle of the merge window, I'll not have time to look at
> this for a few weeks at the earliest, sorry.
>
> thanks,
>
> greg k-h
> --
> To unsubscribe from this list: send the line "unsubscribe
> linux-kernel" in
> the body of a message to [email protected]
> <mailto:[email protected]>
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>

2014-12-24 09:28:19

by Bill Huang

[permalink] [raw]
Subject: Re: [RFC 1/1] driver core: re-order dpm_list after a succussful probe



On 12/23/2014 10:26 PM, Bibek Basu wrote:
> sorry for the delay in responding.
>
> On Thu, Dec 18, 2014 at 1:35 PM, Bill Huang <[email protected]
> <mailto:[email protected]>> wrote:
>
>
>
> On 12/17/2014 10:47 PM, Bibek Basu wrote:
>
> Hi Bill,
>
> Though I like your solution, I have a usecase where the driver probe
> sequence itself is not right. Both the driver are module_init
> but one
> depends on another during suspend sequence.
> In such a situation, my system hangs. What do you suggest to do
> in that
> case? Should I get my driver registration sequence right and how?
> Moving tegra-pcie driver above in the probe sequence by making the
> driver subsystem_initcall solved the issue I am facing with this
> patch.
> But I don't think that's allowed solution?
>
>
> To change the probe sequence, use defer probe is the right choice.
>
> What if there is no direct dependency to defer probe ?

In your example, tegra-pcie driver doesn't have init dependency against
pcieport but do have pm suspend dependency in between the two. Though
I'm not familiar with pcie driver framework but it sounds like pcieport
driver is framework driver so it might not be able to know/expect there
will be "tegra-pcie" driver been probed later, that means defer probe
cannot be used to resolve the pm sequence.

If suspend function of tegra-pcie driver cannot be hooked to pcieport
driver (so it won't hang while pcieport is suspending) and explicitly to
be called to suspend after pcieport driver has done all its access to
the controller then I personally think moving tegra-pcie to
subsystem_initcall is an acceptable fix, but I'm not expert of the pm
design so maybe there is some other proper fix...
>
>
>
> Example:
>
> Probe sequence:
> driver pcieport
> driver tegra-pcie
>
> Due to your patch, suspend_noirq for tegra_pcie will be called
> before
>
>
> Are you sure? My change will only affect pm devices in dpm_list,
> suspend_noirq should still be called after all devices in dpm_list
> were suspended.
>
> Yes, all lists are affected if you update dpm_list. Because using
> dpm_list, dm_prepared_list is prepared . During suspend, using
> dpm_prepared_list, dpm_suspended_list is prepared. And using
> dpc_suspended_list, in dpm_suspend_late, dpm_late_early_list is prepared
> and which is used to call suspend_noirq callbacks and also prepare
> dpm_noirq_list. So all are related!!!
>
True they are all related and if both pcieport and tegra-pcie have
suspend_noirq callback defined (I guess so since you said you have
problem here:), then my fix will change the suspend sequence of the two.

This is my main reason to send this RFC since it could break a lot of
existing working pm sequence.
>
> pcieport. While pcieport tries to read through
> pci_bus_read_config_dword
> with clocks and power off to the pcie controller and eventually
> leads to
> a crash.
>
>
>
> regards
> Bibek
>
> On Fri, Dec 12, 2014 at 9:04 PM, Greg KH
> <[email protected] <mailto:[email protected]>
> <mailto:gregkh@__linuxfoundation.org
> <mailto:[email protected]>>> wrote:
>
> On Fri, Dec 12, 2014 at 03:50:15AM -0800, Bill Huang wrote:
> > The dpm_list was added in the call "device_add" and when
> we do defer
> > probe we'll explicitly move the probed device to be the
> last in the
> > dpm_list, we should do the same for the normal probe
> since there are
> > cases that we won't have chance to do defer probe to
> change the
> PM order
> > as the below example.
> >
> > If we would like the device driver A to be suspended
> earlier than the
> > device driver B, we won't have chance to do defer probe
> to fix the
> > suspend dependency since at the time device driver A is
> probed,
> device B
> > was up and running.
> >
> > Device A was added
> > Device B was added
> > Driver for device B was binded
> > Driver for device A was binded
> >
> > Signed-off-by: Bill Huang <[email protected]
> <mailto:[email protected]>
> <mailto:[email protected] <mailto:[email protected]>>>
>
> > ---
> >
> > It seems to me this is a bug in the core driver, but I'm
> not sure
> how should
> > we fix it.
> >
> > - Do we have better fix?
> > - This proposed fix or any other fix might introduces
> side effect
> that breaks
> > existing working suspend dependencies which happen to work
> based on the
> > existing wrong suspend order.
> >
> > Any thoughts? Thanks.
> >
> > drivers/base/dd.c | 4 ++++
> > 1 file changed, 4 insertions(+)
> >
> > diff --git a/drivers/base/dd.c b/drivers/base/dd.c
> > index cdc779c..54886d2 100644
> > --- a/drivers/base/dd.c
> > +++ b/drivers/base/dd.c
> > @@ -308,6 +308,10 @@ static int really_probe(struct
> device *dev,
> struct device_driver *drv)
> > goto probe_failed;
> > }
> >
> > + device_pm_lock();
> > + device_pm_move_last(dev);
> > + device_pm_unlock();
> > +
> > driver_bound(dev);
> > ret = 1;
> > pr_debug("bus: '%s': %s: bound device %s to driver
> %s\n",
>
>
> Adding Grant, as he did the deferred probe stuff...
>
> And it's the middle of the merge window, I'll not have time
> to look at
> this for a few weeks at the earliest, sorry.
>
> thanks,
>
> greg k-h
> --
> To unsubscribe from this list: send the line "unsubscribe
> linux-kernel" in
> the body of a message to [email protected]
> <mailto:[email protected]>
> <mailto:[email protected].__org
> <mailto:[email protected]>>
> More majordomo info at
> http://vger.kernel.org/__majordomo-info.html
> <http://vger.kernel.org/majordomo-info.html>
> Please read the FAQ at http://www.tux.org/lkml/
>