Re: [PATCH] staging: ram_console: Fix section mismatches

[Stephen Boyd <sboyd@codeaurora.org>]

On 03/08/12 10:43, Greg KH wrote:
> On Thu, Mar 08, 2012 at 10:34:14AM -0800, Stephen Boyd wrote:
>> On 03/08/12 10:23, Greg KH wrote:
>>> On Thu, Mar 08, 2012 at 10:12:07AM -0800, Stephen Boyd wrote:
>>>> On 03/08/12 09:56, Greg KH wrote:
>>>>> On Thu, Mar 08, 2012 at 01:08:04AM -0800, Stephen Boyd wrote:
>>>>>> WARNING: vmlinux.o(.text+0x25d5fc): Section mismatch in reference
>>>>>> from the function ram_console_driver_probe() to the function
>>>>>> .init.text:ram_console_init()
>>>>>> The function ram_console_driver_probe() references
>>>>>> the function __init ram_console_init().
>>>>>> This is often because ram_console_driver_probe lacks a __init
>>>>>> annotation or the annotation of ram_console_init is wrong.
>>>>>>
>>>>>> Mark ram_console_driver_probe() as __devinit because it's a probe
>>>>>> function and propagate the __devinit markings to the __init
>>>>>> functions the probe calls.
>>>>> What .config configuration causes this to happen?  I don't see this here
>>>>> in my builds, what am I doing wrong?
>>>>>
>>>> #
>>>> # Android
>>>> #
>>>> CONFIG_ANDROID=y
>>>> # CONFIG_ANDROID_BINDER_IPC is not set
>>>> # CONFIG_ASHMEM is not set
>>>> # CONFIG_ANDROID_LOGGER is not set
>>>> CONFIG_ANDROID_RAM_CONSOLE=y
>>>> # CONFIG_ANDROID_RAM_CONSOLE_ERROR_CORRECTION is not set
>>>> # CONFIG_ANDROID_TIMED_OUTPUT is not set
>>>> # CONFIG_ANDROID_LOW_MEMORY_KILLER is not set
>>>> # CONFIG_ANDROID_SWITCH is not set
>>>> # CONFIG_PHONE is not set
>>>>
>>>>
>>>> Perhaps you're missing this patch if you're on an ARM compiler?
>>>>
>>>> 6e2e340 (ARM: 7324/1: modpost: Fix section warnings for ARM for many
>>>> compilers, 2012-02-14)
>>> Nope, I'm building this on x86-64 which warns on this type of thing all
>>> the time.
>>>
>>> My .config looks like this:
>>>
>>> #
>>> # Android
>>> #
>>> CONFIG_ANDROID=y
>>> CONFIG_ANDROID_BINDER_IPC=y
>>> CONFIG_ASHMEM=y
>>> CONFIG_ANDROID_LOGGER=m
>>> CONFIG_ANDROID_PERSISTENT_RAM=y
>>> CONFIG_ANDROID_RAM_CONSOLE=y
>>> CONFIG_ANDROID_TIMED_OUTPUT=y
>>> # CONFIG_ANDROID_TIMED_GPIO is not set
>>> CONFIG_ANDROID_LOW_MEMORY_KILLER=y
>>> CONFIG_ANDROID_SWITCH=m
>>> CONFIG_ANDROID_SWITCH_GPIO=m
>>> CONFIG_ANDROID_INTF_ALARM=y
>>> CONFIG_ANDROID_INTF_ALARM_DEV=y
>>> CONFIG_ANDROID_ALARM_OLDDRV_COMPAT=y
>>>
>>> And I can't duplicate this at all.  Could the recent fixes that John
>>> sent me be the reason?  Or something else?
>>>
>> This patch is based on your staging-next branch at c5ee121 (staging:
>> android: ram_console: drop verbose ram_console support, 2012-03-07). It
>> applied that ARM patch on top because I'm compiling with ARM.
> The ram_console just got reworked a bunch by the patches I applied a few
> hours ago (and you had responded to that thread), so even if I wanted to
> take your patch, I can't :)
>
> Care to redo it if it is needed after the rework that just happened?
>

Sure. Now it looks like:

WARNING: vmlinux.o(.text+0xfcf6e): Section mismatch in reference from
the function ram_console_driver_probe() to the function
.init.text:persistent_ram_init_ringbuffer()
The function ram_console_driver_probe() references
the function __init persistent_ram_init_ringbuffer().
This is often because ram_console_driver_probe lacks a __init
annotation or the annotation of persistent_ram_init_ringbuffer is wrong.

I think this wants to use platform_driver_probe() so I've changed the patch.

>> It looks like aggressive inlining by the x86 compiler hides this from
>> you. I see that if I mark ram_console_init() as noinline
>>
>> diff --git a/drivers/staging/android/ram_console.c b/drivers/staging/android/ram_console.c
>> index 73215e2..c468fa2 100644
>> --- a/drivers/staging/android/ram_console.c
>> +++ b/drivers/staging/android/ram_console.c
>> @@ -205,7 +205,7 @@ ram_console_save_old(struct ram_console_buffer *buffer, char *dest)
>>                &buffer->data[0], buffer->start);
>>  }
>>  
>> -static int __init ram_console_init(struct ram_console_buffer *buffer,
>> +static noinline int __init ram_console_init(struct ram_console_buffer *buffer,
>>                                    size_t buffer_size, char *old_buf)
>>  {
>>  #ifdef CONFIG_ANDROID_RAM_CONSOLE_ERROR_CORRECTION
>>
>>
>> then I see the section mismatch when compiling on x86. Otherwise I don't
>> see anything. Is there a bug in the section mismatch detection with
>> respect to compiler inlining?
> Possibly, but probably not, if the code is inlined, then there is no
> error happening, so there's nothing to detect, right?

Presumably the person writing the driver wanted this function thrown
away after kernel init but now it's not being thrown away because it got
inlined into a function that isn't marked correctly. That sounds like
some type of bug to me either in someone's brain or in some random
compiler. But I suppose you're talking about section mismatch detection
having a bug and I agree it's probably impossible to catch this case
when all we have to work with is the compiler's output.

-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.