[PATCH] powerpc/vdso: Separate vvar vma from vdso

[Laurent Dufour]

Hi Christophe and Dimitry,

Le 27/03/2021 à 18:43, Dmitry Safonov a écrit :
> Hi Christophe,
> 
> On 3/27/21 5:19 PM, Christophe Leroy wrote:
> [..]
>>> I opportunistically Cc stable on it: I understand that usually such
>>> stuff isn't a stable material, but that will allow us in CRIU have
>>> one workaround less that is needed just for one release (v5.11) on
>>> one platform (ppc64), which we otherwise have to maintain.
>>
>> Why is that a workaround, and why for one release only ? I think the
>> solution proposed by Laurentto use the aux vector AT_SYSINFO_EHDR should
>> work with any past and future release.
> 
> Yeah, I guess.
> Previously, (before v5.11/power) all kernels had ELF start at "[vdso]"
> VMA start, now we'll have to carry the offset in the VMA. Probably, not
> the worst thing, but as it will be only for v5.11 release it can break,
> so needs separate testing.
> Kinda life was a bit easier without this additional code.
The assumption that ELF header is at the start of "[vdso]" is perhaps not a good 
one, but using a "[vvar]" section looks more conventional and allows to clearly 
identify the data part. I'd argue for this option.

> 
>>> I wouldn't go as far as to say that the commit 511157ab641e is ABI
>>> regression as no other userspace got broken, but I'd really appreciate
>>> if it gets backported to v5.11 after v5.12 is released, so as not
>>> to complicate already non-simple CRIU-vdso code. Thanks!
>>>
>>> Cc: Andrei Vagin <avagin at gmail.com>
>>> Cc: Andy Lutomirski <luto at kernel.org>
>>> Cc: Benjamin Herrenschmidt <benh at kernel.crashing.org>
>>> Cc: Christophe Leroy <christophe.leroy at csgroup.eu>
>>> Cc: Laurent Dufour <ldufour at linux.ibm.com>
>>> Cc: Michael Ellerman <mpe at ellerman.id.au>
>>> Cc: Paul Mackerras <paulus at samba.org>
>>> Cc: linuxppc-dev at lists.ozlabs.org
>>> Cc: stable at vger.kernel.org # v5.11
>>> [1]: https://github.com/checkpoint-restore/criu/issues/1417
>>> Signed-off-by: Dmitry Safonov <dima at arista.com>
>>> Tested-by: Christophe Leroy <christophe.leroy at csgroup.eu>
>>
>> I tested it with sifreturn_vdso selftest and it worked, because that
>> selftest doesn't involve VDSO data.
> 
> Thanks again on helping with testing it, I appreciate it!
> 
>> But if I do a mremap() on the VDSO text vma without remapping VVAR to
>> keep the same distance between the two vmas, gettimeofday() crashes. The
>> reason is that the code obtains the address of the data by calculating a
>> fix difference from its own address with the below macro, the delta
>> being resolved at link time:
>>
>> .macro get_datapage ptr
>>      bcl    20, 31, .+4
>> 999:
>>      mflr    \ptr
>> #if CONFIG_PPC_PAGE_SHIFT > 14
>>      addis    \ptr, \ptr, (_vdso_datapage - 999b)@ha
>> #endif
>>      addi    \ptr, \ptr, (_vdso_datapage - 999b)@l
>> .endm
>>
>> So the datapage needs to remain at the same distance from the code at
>> all time.
>>
>> Wondering how the other architectures do to have two independent VMAs
>> and be able to move one independently of the other.
> 
> It's alright as far as I know. If userspace remaps vdso/vvar it should
> be aware of this (CRIU keeps this in mind, also old vdso image is dumped
> to compare on restore with the one that the host has).

I do agree, playing with the VDSO mapping needs the application to be aware of 
the mapping details, and prior to 83d3f0e90c6c "powerpc/mm: tracking vDSO 
remap", remapping the VDSO was not working on PowerPC and nobody complained...

Laurent.