[PATCH] ARM: pxa: Fix pxa3xx-u2d crash when ULPI not used

[Igor Grinberg]

 On 09/05/10 13:43, Eric Miao wrote:
> On Sun, Sep 5, 2010 at 4:35 PM, Igor Grinberg <grinberg at compulab.co.il> wrote:
>>  On 09/05/10 11:25, Marek Vasut wrote:
>>> Dne Ne 5. září 2010 10:16:48 Igor Grinberg napsal(a):
>>>>  On 09/05/10 11:01, Marek Vasut wrote:
>>>>> Dne Ne 5. září 2010 09:54:31 Igor Grinberg napsal(a):
>>>>>>  On 09/03/10 23:35, Marek Vasut wrote:
>>>>>>> In case the pxa3xx-u2d driver isn't used, probing of ohci-pxa27x will
>>>>>>> cause an ugly kernel crash (NULL pointer dereference in
>>>>>>> pxa3xx_u2d_start_hc(), because struct u2d is NULL and clk_enable() call
>>>>>>> will crash the kernel, trying to access it).
>>>>>> ohci code checks for pxa3xx cpu and only then runs start/stop hc.
>>>>> Exactly ... and in case "struct pxa3xx_u2d_ulpi *u2d" is NULL, clk_enable
>>>>> will crash the kernel.
>>>>>
>>>>>> pxa3xx_ulpi.c is compiled if CONFIG_PXA3xx is selected.
>>>>>> The device <-> driver binding should not be a problem, so the
>>>>>> pxa3xx_u2d_probe() will run.
>>>>>> The only case, I see, when struct u2d does not exist is failure of the
>>>>>> probe function. If this is the case, we are having an abnormal execution
>>>>>> and if your patch is dealing with this issue, shouldn't you comment it
>>>>>> as such?
>>>>> Not at all ... if the pxa3xx-u2d driver isn't loaded at all, the function
>>>>> (start/stop hc) is still called, but struct pxa3xx_u2d_ulpi *u2d is NULL.
>>>>> In this case, if you call clk_enable(u2d->clk), you crash the kernel
>>>>> (because u2d is NULL).
>>>> How, can it happen, that "pxa3xx-u2d driver isn't loaded at all"?
>>>> This can happen only if you rip out the device registration or hack a
>>>> Makefile. I don't see any other way... is there?
>>> If you don't call pxa3xx_set_u2d_info() ?
>> Oh... right.
>> I've added it this way, so boards can control u2d existence and forgot it is there...
>> Buggy me... :(
>>
> Igor,
>
> So do you this as a proper fix, or there is better way out?

If we want to keep it the most straight-forward way so it is fine.

Another possible ways would be:

1) create a new flag, lets say PORT2_USE_U2DC in pxaohci_platform_data.
This is a relatively clean way of making ohci aware of
u2d existence at runtime and eliminates the calls to functions
of non-existing (not loaded) driver...

2) use something like:
struct u2d_hc_ops {
    int (*start_hc)(...);
    void (*stop_hc)(...);
}

in board init code register the ops structure via the pxa_ohci platform data.

This way adds some more pollution to the pxa_ohci glue with code
relevant only to pxa3xx, achieving the same as the 1st way, but also
can be useful for u2dc otg/peripheral driver (if it will come some day...).

I'm fine with both (Marek's patch or my proposal), so tell me what
looks better to you.

-- 
Regards,
Igor.