[modules,0/2] Fix handling of exit unwinding sections (on ARM)

Message

Matthias Schiffer June 3, 2019, 10:57 a.m. UTC

For some time (050d18d1c651 "ARM: 8650/1: module: handle negative
R_ARM_PREL31 addends correctly", v4.11+), building a kernel without
CONFIG_MODULE_UNLOAD would lead to module loads failing on ARM systems with
certain memory layouts, with messages like:

  imx_sdma: section 16 reloc 0 sym '': relocation 42 out of range
  (0x7f015260 -> 0xc0f5a5e8)

(0x7f015260 is in the module load area, 0xc0f5a5e8 a regular vmalloc
address; relocation 42 is R_ARM_PREL31)

This is caused by relocatiosn in the .ARM.extab.exit.text and
.ARM.exidx.exit.text sections referencing the .exit.text section. As the
module loader will omit loading .exit.text without CONFIG_MODULE_UNLOAD,
there will be relocations from loaded to unloaded sections; the resulting
huge offsets trigger the sanity checks added in 050d18d1c651.

IA64 might be affected by a similar issue - sections with names like
.IA_64.unwind.exit.text and .IA_64.unwind_info.exit.text appear in the ld
script - but I don't know much about that arch.

Also, I'm not sure if this is stable-worthy - just enabling
CONFIG_MODULE_UNLOAD should be a viable workaround on affected kernels.


Kind regards,
Matthias


Matthias Schiffer (2):
  module: allow arch overrides for .exit section names
  ARM: module: recognize unwind exit sections

 arch/arm/include/asm/module.h | 11 +++++++++++
 include/linux/moduleloader.h  |  8 ++++++++
 kernel/module.c               |  2 +-
 3 files changed, 20 insertions(+), 1 deletion(-)

Comments

Matthias Schiffer June 6, 2019, 8:14 a.m. UTC | #1

On Mon, 2019-06-03 at 12:57 +0200, Matthias Schiffer wrote:
> For some time (050d18d1c651 "ARM: 8650/1: module: handle negative
> R_ARM_PREL31 addends correctly", v4.11+), building a kernel without
> CONFIG_MODULE_UNLOAD would lead to module loads failing on ARM
> systems with
> certain memory layouts, with messages like:
> 
>   imx_sdma: section 16 reloc 0 sym '': relocation 42 out of range
>   (0x7f015260 -> 0xc0f5a5e8)
> 
> (0x7f015260 is in the module load area, 0xc0f5a5e8 a regular vmalloc
> address; relocation 42 is R_ARM_PREL31)
> 
> This is caused by relocatiosn in the .ARM.extab.exit.text and
> .ARM.exidx.exit.text sections referencing the .exit.text section. As
> the
> module loader will omit loading .exit.text without
> CONFIG_MODULE_UNLOAD,
> there will be relocations from loaded to unloaded sections; the
> resulting
> huge offsets trigger the sanity checks added in 050d18d1c651.
> 
> IA64 might be affected by a similar issue - sections with names like
> .IA_64.unwind.exit.text and .IA_64.unwind_info.exit.text appear in
> the ld
> script - but I don't know much about that arch.
> 
> Also, I'm not sure if this is stable-worthy - just enabling
> CONFIG_MODULE_UNLOAD should be a viable workaround on affected
> kernels.
> 
> 
> Kind regards,
> Matthias


Hi,
any comments on these patches? If not, who is going to take them in
their tree?

Note that I'll be out of office for the next week, and I won't be able
to read my mail during this period.


Kind regards,
Matthias




> 
> 
> Matthias Schiffer (2):
>   module: allow arch overrides for .exit section names
>   ARM: module: recognize unwind exit sections
> 
>  arch/arm/include/asm/module.h | 11 +++++++++++
>  include/linux/moduleloader.h  |  8 ++++++++
>  kernel/module.c               |  2 +-
>  3 files changed, 20 insertions(+), 1 deletion(-)
>

Jessica Yu June 6, 2019, 3:21 p.m. UTC | #2

+++ Matthias Schiffer [06/06/19 10:14 +0200]:
>On Mon, 2019-06-03 at 12:57 +0200, Matthias Schiffer wrote:
>> For some time (050d18d1c651 "ARM: 8650/1: module: handle negative
>> R_ARM_PREL31 addends correctly", v4.11+), building a kernel without
>> CONFIG_MODULE_UNLOAD would lead to module loads failing on ARM
>> systems with
>> certain memory layouts, with messages like:
>>
>>   imx_sdma: section 16 reloc 0 sym '': relocation 42 out of range
>>   (0x7f015260 -> 0xc0f5a5e8)
>>
>> (0x7f015260 is in the module load area, 0xc0f5a5e8 a regular vmalloc
>> address; relocation 42 is R_ARM_PREL31)
>>
>> This is caused by relocatiosn in the .ARM.extab.exit.text and
>> .ARM.exidx.exit.text sections referencing the .exit.text section. As
>> the
>> module loader will omit loading .exit.text without
>> CONFIG_MODULE_UNLOAD,
>> there will be relocations from loaded to unloaded sections; the
>> resulting
>> huge offsets trigger the sanity checks added in 050d18d1c651.
>>
>> IA64 might be affected by a similar issue - sections with names like
>> .IA_64.unwind.exit.text and .IA_64.unwind_info.exit.text appear in
>> the ld
>> script - but I don't know much about that arch.
>>
>> Also, I'm not sure if this is stable-worthy - just enabling
>> CONFIG_MODULE_UNLOAD should be a viable workaround on affected
>> kernels.
>>
>>
>> Kind regards,
>> Matthias
>
>
>Hi,
>any comments on these patches? If not, who is going to take them in
>their tree?

I don't mind either way. I can take the patches through my tree if
Russell ack's the second one (after comments have been addressed).

Thanks!

Jessica