Daniel Cashman <[email protected]> writes:
> On 10/28/2015 08:41 PM, Eric W. Biederman wrote:
>> Dan Cashman <[email protected]> writes:
>>
>>>>> This all would be much cleaner if the arm architecture code were just to
>>>>> register the sysctl itself.
>>>>>
>>>>> As it sits this looks like a patchset that does not meaninfully bisect,
>>>>> and would result in code that is hard to trace and understand.
>>>>
>>>> I believe the intent is to follow up with more architecture specific
>>>> patches to allow each architecture to define the number of bits to use
>>>
>>> Yes. I included these patches together because they provide mutual
>>> context, but each has a different outcome and they could be taken
>>> separately.
>>
>> They can not. The first patch is incomplete by itself.
>
> Could you be more specific in what makes the first patch incomplete? Is
> it because it is essentially a no-op without additional architecture
> changes (e.g. the second patch) or is it specifically because it
> introduces and uses the three "mmap_rnd_bits*" variables without
> defining them? If the former, I'd like to avoid combining the general
> procfs change with any architecture-specific one(s). If the latter, I
> hope the proposal below addresses that.
A bit of both. The fact that the code can not compile in the first
patch because of missing variables is distressing. Having the arch
specific code as a separate patch is fine, but they need to remain in
the same patchset.
>>> The arm architecture-specific portion allows the changing
>>> of the number of bits used for mmap ASLR, useful even without the
>>> sysctl. The sysctl patch (patch 1) provides another way of setting
>>> this value, and the hope is that this will be adopted across multiple
>>> architectures, with the arm changes (patch 2) providing an example. I
>>> hope to follow this with changes to arm64 and x86, for example.
>>
>> If you want to make the code generic. Please maximize the sharing.
>> That is please define the variables in a generic location, as well
>> as the Kconfig variables (if possible).
>>
>> As it is you have an architecture specific piece of code that can not be
>> reused without duplicating code, and that is just begging for problems.
>
> I think it would make sense to move the variable definitions into
> mm/mmap.c, included conditionally based on the presence of
> CONFIG_ARCH_MMAP_RND_BITS.
>
> As for the Kconfigs, I am open to suggestions. I considered declaring
> and documenting ARCH_MMAP_RND_BITS in arch/Kconfig, but I would like it
> to be bounded in range by the _MIN and _MAX values, which necessarily
> must be defined in the arch-specific Kconfigs. Thus, we'd have
> ARCH_MMAP_RND_BITS declared in arch/Kconfig as it currently is in
> arch/arm/Kconfig defaulting to _MIN, and would declare both the _MIN and
> _MAX in arch/Kconfig, while specifying default values in
> arch/${ARCH}/Kconfig.
>
> Would these changes be more acceptable?
Yes. I don't think you can do much about the Kconfigs so I would not
worry about that too much.
Eric
On 11/01/2015 01:50 PM, Eric W. Biederman wrote:
> Daniel Cashman <[email protected]> writes:
>
>> On 10/28/2015 08:41 PM, Eric W. Biederman wrote:
>>> Dan Cashman <[email protected]> writes:
>>>
>>>>>> This all would be much cleaner if the arm architecture code were just to
>>>>>> register the sysctl itself.
>>>>>>
>>>>>> As it sits this looks like a patchset that does not meaninfully bisect,
>>>>>> and would result in code that is hard to trace and understand.
>>>>>
>>>>> I believe the intent is to follow up with more architecture specific
>>>>> patches to allow each architecture to define the number of bits to use
>>>>
>>>> Yes. I included these patches together because they provide mutual
>>>> context, but each has a different outcome and they could be taken
>>>> separately.
>>>
>>> They can not. The first patch is incomplete by itself.
>>
>> Could you be more specific in what makes the first patch incomplete? Is
>> it because it is essentially a no-op without additional architecture
>> changes (e.g. the second patch) or is it specifically because it
>> introduces and uses the three "mmap_rnd_bits*" variables without
>> defining them? If the former, I'd like to avoid combining the general
>> procfs change with any architecture-specific one(s). If the latter, I
>> hope the proposal below addresses that.
>
> A bit of both. The fact that the code can not compile in the first
> patch because of missing variables is distressing. Having the arch
> specific code as a separate patch is fine, but they need to remain in
> the same patchset.
>
The first patch would compile as long as CONFIG_ARCH_MMAP_RND_BITS were
not defined without also defining the missing variables. I actually
viewed this as a safeguard against attempting to use those variables
without architecture support, but am ok with changing it.
I've gone ahead and submitted [PATCH v2] which aims to reduce
duplication in the arch-specific config files and concerning those
variables. The only caveat is that now the second patch depends on the
first, whereas before it did not.
Thank You,
Dan