Re: [PATCH 3/4] x86,mm: Improve _install_special_mapping and fix x86 vdso naming

[Andy Lutomirski <luto@amacapital.net>]

On Tue, May 20, 2014 at 10:47 AM, Cyrill Gorcunov <gorcunov@gmail.com> wrote:
> On Tue, May 20, 2014 at 10:24:49AM -0700, Andy Lutomirski wrote:
>> On Tue, May 20, 2014 at 10:21 AM, Cyrill Gorcunov <gorcunov@gmail.com> wrote:
>> > On Mon, May 19, 2014 at 03:58:33PM -0700, Andy Lutomirski wrote:
>> >> Using arch_vma_name to give special mappings a name is awkward.  x86
>> >> currently implements it by comparing the start address of the vma to
>> >> the expected address of the vdso.  This requires tracking the start
>> >> address of special mappings and is probably buggy if a special vma
>> >> is split or moved.
>> >>
>> >> Improve _install_special_mapping to just name the vma directly.  Use
>> >> it to give the x86 vvar area a name, which should make CRIU's life
>> >> easier.
>> >>
>> >> As a side effect, the vvar area will show up in core dumps.  This
>> >> could be considered weird and is fixable.  Thoughts?
>> >>
>> >> Cc: Cyrill Gorcunov <gorcunov@openvz.org>
>> >> Cc: Pavel Emelyanov <xemul@parallels.com>
>> >> Signed-off-by: Andy Lutomirski <luto@amacapital.net>
>> >
>> > Hi Andy, thanks a lot for this! I must confess I don't yet know how
>> > would we deal with compat tasks but this is 'must have' mark which
>> > allow us to detect vvar area!
>>
>> Out of curiosity, how does CRIU currently handle checkpointing a
>> restored task?  In current kernels, the "[vdso]" name in maps goes
>> away after mremapping the vdso.
>
>   We use not only [vdso] mark to detect vdso area but also page frame
> number of the living vdso. If mark is not present in procfs output
> we examinate executable areas and check if pfn == vdso_pfn, it's
> a slow path because there migh be a bunch of executable areas and
> touching every of it is not that fast thing, but we simply have no
> choise.

This patch should fix this issue, at least.  If there's still a way to
get a native vdso that doesn't say "[vdso]", please let me know/

>
>   The situation get worse when task was dumped on one kernel and
> then restored on another kernel where vdso content is different
> from one save in image -- is such case as I mentioned we need
> that named vdso proxy which redirect calls to vdso of the system
> where task is restoring. And when such "restored" task get checkpointed
> second time we don't dump new living vdso but save only old vdso
> proxy on disk (detecting it is a different story, in short we
> inject a unique mark into elf header).

Yuck.  But I don't know whether the kernel can help much here.

>
>>
>> I suspect that you'll need kernel changes for compat tasks, since I
>> think that mremapping the vdso on any reasonably modern hardware in a
>> 32-bit task will cause sigreturn to blow up.  This could be fixed by
>> making mremap magical, although adding a new prctl or arch_prctl to
>> reliably move the vdso might be a better bet.
>
> Well, as far as I understand compat code uses abs addressing for
> vvar data and if vvar data position doesn't change we're safe,
> but same time because vvar addresses are not abi I fear one day
> we indeed hit the problems and the only solution would be
> to use kernel's help. But again, Andy, I didn't think much
> about implementing compat mode in criu yet so i might be
> missing some details.

Prior to 3.15, the compat code didn't have vvar data at all.  In 3.15
and up, the vvar data is accessed using PC-relative addressing, even
in compat mode (using the usual call; mov trick to read EIP).

--Andy