Re: [PATCH 2/2] kernel: Support compiling out the prctl syscall

[Josh Triplett]

On November 8, 2016 4:56:54 PM PST, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>On Tue, Nov 8, 2016 at 4:47 PM, Josh Triplett <josh@xxxxxxxxxxxxxxxx>
>wrote:
>> On Tue, Nov 08, 2016 at 04:40:02PM -0800, Kees Cook wrote:
>>> On Tue, Nov 8, 2016 at 4:18 PM, Josh Triplett
><josh@xxxxxxxxxxxxxxxx> wrote:
>>> > Some embedded systems can do without the prctl syscall, saving
>some
>>> > space.
>>> >
>>> > This also avoids regular increases in tinyconfig size as people
>add more
>>> > non-optional functionality to prctl (observed via the 0-day kernel
>>> > infrastructure).
>>> >
>>> > bloat-o-meter results:
>>> >
>>> > add/remove: 0/3 grow/shrink: 0/1 up/down: 0/-2143 (-2143)
>>> > function                                     old     new   delta
>>> > offsets                                       23      12     -11
>>> > prctl_set_auxv                                97       -     -97
>>> > sys_prctl                                    794       -    -794
>>> > prctl_set_mm                                1241       -   -1241
>>> > Total: Before=1902583, After=1900440, chg -0.11%
>>> >
>>> > Signed-off-by: Josh Triplett <josh@xxxxxxxxxxxxxxxx>
>>>
>>> I'm absolutely a fan of doing this, but I wonder how this interacts
>>> with the LSMs that define prctl hooks, etc. I wouldn't expect a
>system
>>> that didn't want prctl to want an LSM, but maybe the LSMs all need
>to
>>> depend on CONFIG_PRCTL now?
>>
>> I did think about that (as well as SECCOMP), but I did confirm that
>the
>> kernel builds fine with allyesconfig minus CONFIG_PRCTL.  An LSM that
>> wants to restrict access to some prctls should be fine with no
>process
>> having any access to prctl. :)  Beyond that, anything wanting
>> configuration via LSM (such as SECCOMP) still exists and functions,
>even
>> if you can't access it from outside the kernel.
>
>Okay, testing that is good, thanks.
>
>Seccomp can use the seccomp() syscall, so missing prctl isn't a big
>deal there.
>
>Things like Yama, though, are almost useless in the !PRCTL case. I
>think a "depends on PRCTL" should be added at least to Yama. All the
>other LSMs are configured in other ways, and they'll just have some
>dead code around their prctl hooks; no big deal.

OK, I'll add that dependency to Yama in v2.

>This does also beg the question about how to configure some process
>behaviors by default if PRCTL is disabled, but if people want those
>things, they can write patches, I would think. :)

Agreed.