lkml.org 
[lkml]   [2014]   [Aug]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: + prctl-pr_set_mm-introduce-pr_set_mm_map-operation-v3.patch added to -mm tree
On 08/23, Cyrill Gorcunov wrote:
>
> On Sat, Aug 23, 2014 at 05:32:22PM +0200, Oleg Nesterov wrote:
> >
> > And btw, where do you see RLIMIT_STACK in do_shmat() ?
>
> Indirectly, though start_stack pointer. We assign it in setup_arg_pages taking into
> account RLIMIT_STACK value, then do_shmat operates with start_stack pointer,
> thus when we setup some new value into start_stack we at least should
> not exceed old RLIMIT_STACK?

Still can't understand how do_shmat() connects to RLIMIT_STACK, even
indirectly. setup_arg_pages() obviously uses RLIMIT_STACK, but with or
without the patch I sent do_shmat() behaviour does not depend on
RLIMIT_STACK at all.

> > > arg_start, arg_end, env_start, env_end
> > > really point to existing VMAs, strictly speaking the probgram can unmap
> > > all own VMAs except executable one and continue running without problem
> > > but this is not that practical I think and at first iteration I prefer
> > > more severe tests here on VMAs
> >
> > But, again, for what? There are only used to report this info via /proc/.
>
> Because it's close to how loading elf proceeds. This prctl is like emulating
> micro-elf loader ;) Indeed there is no problem for kernel if all these members
> are pointing to some already umapped VMAs,

Yes. And whatever checks you add into prctl(), an application can simply
do mmap() before prctl() just to fool it, and unmap (say) arg_start right
after that.

> but earlier we required cap-sysadmin
> to be grated so that if someone calls for this prctl he must be knowing what
> he is doing, and if some other unpredicted change in kernel code cause this
> prctl to panic the kernel such action could be initiated by sysadmin only,
> not regular user. Now I've released the security requirement so I thought
> let be more paranoid and require all VMAs to present, all rlimits to be
> in good shape and such even if the members we're modifying are used for
> procfs statistics output mostly, we always have a way to relax this tests
> but not the reverse. That was the main idea ;)

Following this logic you should also verify that KSTK_ESP(group_leader)
falls into ->start_stack area or return an error otherwise.

As for elf loader, of course it sets ->start_stack/etc correctly and they
can't point to nowehere, but this is only because it actually creates these
segments. And I guess the c/r tools should do this "correctly" too, but
prctl() itself can never verify this.

OK, lets look at this from another angle. Suppose that an application switches
to another stack and unmaps ->start_stack area. Now, how are you going to
restore it correctly if ->start_stack points to nowhere? Are you going
to "fix" ->start_stack? I do not think this would be correct. Or, are you
going to create the additional vma just to make prctl() happy?

Oleg.



\
 
 \ /
  Last update: 2014-08-23 21:41    [W:0.046 / U:0.180 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site