Re: [PATCH 0/2] net: Fix crashes due to activity during suspend

[Florian Fainelli]

On 08/23/2017 10:13 AM, Florian Fainelli wrote:
> On 08/23/2017 04:45 AM, Geert Uytterhoeven wrote:
>> Hi Florian,
>>
>> On Tue, Aug 22, 2017 at 8:49 PM, Florian Fainelli <f.fainelli@xxxxxxxxx> wrote:
>>> On 08/22/2017 11:37 AM, Geert Uytterhoeven wrote:
>>>> If an Ethernet device is used while the device is suspended, the system may
>>>> crash.
>>>>
>>>> E.g. on sh73a0/kzm9g and r8a73a4/ape6evm, the external Ethernet chip is
>>>> driven by a PM controlled clock.  If the Ethernet registers are accessed
>>>> while the clock is not running, the system will crash with an imprecise
>>>> external abort.
>>>>
>>>> This patch series fixes two of such crashes:
>>>>   1. The first patch prevents the PHY polling state machine from accessing
>>>>      PHY registers while a device is suspended,
>>>>   2. The second patch prevents the net core from trying to transmit packets
>>>>      when an smsc911x device is suspended.
>>>>
>>>> Both crashes can be reproduced on sh73a0/kzm9g and r8a73a4/ape6evm during
>>>> s2ram (rarely), or by using pm_test (more likely to trigger):
>>>>
>>>>     # echo 0 > /sys/module/printk/parameters/console_suspend
>>>>     # echo platform > /sys/power/pm_test
>>>>     # echo mem > /sys/power/state
>>>>
>>>> With this series applied, my test systems survive a loop of 100 test
>>>> suspends.
>>>
>>> It seems to me like part, if not the entire problem is that smsc91xx's
>>> suspend and resume functions are way too simplistic and absolutely do
>>> not manage the PHY during suspend/resume, the PHY state machine is not
>>> even stopped, so of course, this will cause bus errors if you access
>>> those registers.
>>>
>>> You are addressing this as part of patch 2, but this seems to me like
>>> this is still a bit incomplete and you'd need at least phy_stop() and/or
>>> phy_suspend() (does a power down of the PHY) and phy_start() and/or
>>> phy_resume() calls to complete the PHY state machine shutdown during
>>> suspend.
>>>
>>> Have you tried that?
>>
>> Unfortunately that doesn't help.
>> In state PHY_HALTED, the PHY state machine still calls the .adjust_link()
>> callback while the device is suspended.
> 
> Humm that is correct yes.
> 
>>
>> Do you have a clue? This is too far beyond my phy-foo...
> 
> I was initially contemplating a revert of
> 7ad813f208533cebfcc32d3d7474dc1677d1b09a ("net: phy: Correctly process
> PHY_HALTED in phy_stop_machine()") but this is not the root of the
> problem. The problem really is that phy_stop() does not wait for the PHY
> state machine to be stopped so you cannot rely on that and past the
> function return be offered any guarantees that adjust_link is not called.
> 
> We seem to be getting away with that in most drivers because when we see
> phydev->link = 0, we either do nothing or actually turn of the HW block.
> 
> How about we export phy_stop_machine() to drivers which would provide a
> synchronization point that would ensure that no HW accesses are done
> past this point?
> 
> I am absolutely not clear on the implications of using a freezable
> workqueue with respect to the PHY state machine and how devices are
> going to wind-up being powered down or not...

Geert, as you may have notice a revert of the change was sent so 4.13
should be fine, but ultimately I would like to put the non-reverted code
back in after we add a few safeguards:

- David Daney reported he could crash the kernel by calling just
phy_disconnect() with no prior phy_stop() which is not "legal" but
should not crash either

- and you reported the bus errors on smsc911x when we call adjust_link
during suspend, and due to a lack of hard synchronization so phy_stop()
here does not give you enough guarantees to let you turn off power to
the smsc911x block

If that seems accurate then we can work on something that should be
working again (famous last words).

Thanks
-- 
Florian