Re: question on [uprobes] special vma

[Jeff Xu <jeffxu@chromium.org>]

Hi Oleg

On Fri, Aug 2, 2024 at 2:00 AM Oleg Nesterov <oleg@redhat.com> wrote:
>
> Hi Jeff,
>
> On 08/01, Jeff Xu wrote:
> >
> > __create_xol_area() calls  _install_special_mapping() to create a vma
> > named [uprobes].
> >
> > I'm trying to find out the lifetime of this uprobes vma, e.g. when it
> > is created, will it ever be unmapped/remapped/changed during the
> > lifetime of the process.
> >
> > If  the uprobes vma remains the same during the lifetime of the
> > process,
>
> Yes,
>
That is nice.

> > I can call mseal on it so user space can't change it, i.e.
> > blocking munmap/mremap/mprotect/mmap, etc.
>
> I didn't even know about mm/mseal.c...
>
The TLDR: once the memory area is sealed,
mmap/munmap/mremap/mprotect/some of madvise will be blocked.

https://docs.kernel.org/userspace-api/mseal.html

On a related topic, Adhemerval Zanella has an RFC adding mseal support
for dynamic linker.

https://public-inbox.org/libc-alpha/20240731200307.2269811-1-adhemerval.zanella@linaro.org/

> at first glance do_mseal() just adds VM_SEALED for can_modify_vma().
>
> So it seems that xol_add_vma() can just pass the additional VM_SEALED
> flag to _install_special_mapping(), no?
>
Yes. Before I make that change, I want to confirm that [uprobes] VMA
is immutable during the lifetime of the process, hence this email.

> But why it depends on CONFIG_64_BIT?
>
It needn't, but the vm_flags_t in the vma struct for 32 bit is full,
adding 32 bit support will make split/merge code more complex, (at one
point of RFC process, 32 bit is included but later removed),

If we like to have 32 bit support, perhaps we need to refactor the mm
code to expand vm_flags first.

Thanks for your help !
-Jeff

> Oleg.
>