Hello,
On Wed, 26 Aug 2020 at 15:39, Michael Ellerman <mpe@xxxxxxxxxxxxxx> wrote:
> Christophe Leroy <christophe.leroy@xxxxxxxxxx> writes:
[..]
> > arch_remap() gets replaced by vdso_remap()
> >
> > For arch_unmap(), I'm wondering how/what other architectures do, because
> > powerpc seems to be the only one to erase the vdso context pointer when
> > unmapping the vdso.
>
> Yeah. The original unmap/remap stuff was added for CRIU, which I thought
> people tested on other architectures (more than powerpc even).
>
> Possibly no one really cares about vdso unmap though, vs just moving the
> vdso.
>
> We added a test for vdso unmap recently because it happened to trigger a
> KAUP failure, and someone actually hit it & reported it.
You right, CRIU cares much more about moving vDSO.
It's done for each restoree and as on most setups vDSO is premapped and
used by the application - it's actively tested.
Speaking about vDSO unmap - that's concerning only for heterogeneous C/R,
i.e when an application is migrated from a system that uses vDSO to the one
which doesn't - it's much rare scenario.
(for arm it's !CONFIG_VDSO, for x86 it's `vdso=0` boot parameter)
Looking at the code, it seems quite easy to provide/maintain .close() for
vm_special_mapping. A bit harder to add a test from CRIU side
(as glibc won't know on restore that it can't use vdso anymore),
but totally not impossible.
> Running that test on arm64 segfaults:
>
> # ./sigreturn_vdso
> VDSO is at 0xffff8191f000-0xffff8191ffff (4096 bytes)
> Signal delivered OK with VDSO mapped
> VDSO moved to 0xffff8191a000-0xffff8191afff (4096 bytes)
> Signal delivered OK with VDSO moved
> Unmapped VDSO
> Remapped the stack executable
> [ 48.556191] potentially unexpected fatal signal 11.
> [ 48.556752] CPU: 0 PID: 140 Comm: sigreturn_vdso Not tainted 5.9.0-rc2-00057-g2ac69819ba9e #190
> [ 48.556990] Hardware name: linux,dummy-virt (DT)
> [ 48.557336] pstate: 60001000 (nZCv daif -PAN -UAO BTYPE=--)
> [ 48.557475] pc : 0000ffff8191a7bc
> [ 48.557603] lr : 0000ffff8191a7bc
> [ 48.557697] sp : 0000ffffc13c9e90
> [ 48.557873] x29: 0000ffffc13cb0e0 x28: 0000000000000000
> [ 48.558201] x27: 0000000000000000 x26: 0000000000000000
> [ 48.558337] x25: 0000000000000000 x24: 0000000000000000
> [ 48.558754] x23: 0000000000000000 x22: 0000000000000000
> [ 48.558893] x21: 00000000004009b0 x20: 0000000000000000
> [ 48.559046] x19: 0000000000400ff0 x18: 0000000000000000
> [ 48.559180] x17: 0000ffff817da300 x16: 0000000000412010
> [ 48.559312] x15: 0000000000000000 x14: 000000000000001c
> [ 48.559443] x13: 656c626174756365 x12: 7865206b63617473
> [ 48.559625] x11: 0000000000000003 x10: 0101010101010101
> [ 48.559828] x9 : 0000ffff818afda8 x8 : 0000000000000081
> [ 48.559973] x7 : 6174732065687420 x6 : 64657070616d6552
> [ 48.560115] x5 : 000000000e0388bd x4 : 000000000040135d
> [ 48.560270] x3 : 0000000000000000 x2 : 0000000000000001
> [ 48.560412] x1 : 0000000000000003 x0 : 00000000004120b8
> Segmentation fault
> #
>
> So I think we need to keep the unmap hook. Maybe it should be handled by
> the special_mapping stuff generically.
I'll cook a patch for vm_special_mapping if you don't mind :-)
Thanks,
Dmitry
