Re: [PATCH] arch/arm64/kernel/traps: Use find_vma_intersection() in traps for setting si_code

From: Catalin Marinas
Date: Tue Apr 13 2021 - 14:00:41 EST


On Tue, Apr 13, 2021 at 04:52:34PM +0000, Liam Howlett wrote:
> * Catalin Marinas <catalin.marinas@xxxxxxx> [210412 13:44]:
> > On Wed, Apr 07, 2021 at 03:11:06PM +0000, Liam Howlett wrote:
> > > find_vma() will continue to search upwards until the end of the virtual
> > > memory space. This means the si_code would almost never be set to
> > > SEGV_MAPERR even when the address falls outside of any VMA. The result
> > > is that the si_code is not reliable as it may or may not be set to the
> > > correct result, depending on where the address falls in the address
> > > space.
> > >
> > > Using find_vma_intersection() allows for what is intended by only
> > > returning a VMA if it falls within the range provided, in this case a
> > > window of 1.
> > >
> > > Signed-off-by: Liam R. Howlett <Liam.Howlett@xxxxxxxxxx>
> > > ---
> > > arch/arm64/kernel/traps.c | 3 ++-
> > > 1 file changed, 2 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c
> > > index a05d34f0e82a..a44007904a64 100644
> > > --- a/arch/arm64/kernel/traps.c
> > > +++ b/arch/arm64/kernel/traps.c
> > > @@ -383,9 +383,10 @@ void force_signal_inject(int signal, int code, unsigned long address, unsigned i
> > > void arm64_notify_segfault(unsigned long addr)
> > > {
> > > int code;
> > > + unsigned long ut_addr = untagged_addr(addr);
> > >
> > > mmap_read_lock(current->mm);
> > > - if (find_vma(current->mm, untagged_addr(addr)) == NULL)
> > > + if (find_vma_intersection(current->mm, ut_addr, ut_addr + 1) == NULL)
> > > code = SEGV_MAPERR;
> > > else
> > > code = SEGV_ACCERR;
[...]
> > I don't think your change is entirely correct either. We can have a
> > fault below the vma of a stack (with VM_GROWSDOWN) and
> > find_vma_intersection() would return NULL but it should be a SEGV_ACCERR
> > instead.
>
> I'm pretty sure I am missing something. From what you said above, I
> think this means that there can be a user cache fault below the stack
> which should notify the user application that they are not allowed to
> expand the stack by sending a SIGV_ACCERR in the si_code? Is this
> expected behaviour or am I missing a code path to this function?

My point was that find_vma() may return a valid vma where addr < vm_end
but also addr < vm_addr. It's the responsibility of the caller to check
that that vma can be expanded (VM_GROWSDOWN) and we do something like
this in __do_page_fault(). find_vma_intersection(), OTOH, requires addr
>= vm_start.

If we hit this case (addr < vm_start), normally we'd first need to check
whether it's expandable and, if not, return MAPERR. If it's expandable,
it should be ACCERR since something else caused the fault.

Now, I think at least for user_cache_maint_handler(), we can assume that
__do_page_fault() handled any expansion already, so we don't need to
check it here. In this case, your find_vma_intersection() check should
work.

Are there other cases where we invoke arm64_notify_segfault() without a
prior fault? I think in swp_handler() we can bail out early before we
even attempted the access so we may report MAPERR but ACCERR is a better
indication. Also in sys_rt_sigreturn() we always call it as
arm64_notify_segfault(regs->sp). I'm not sure that's correct in all
cases, see restore_altstack().

I guess this code needs some tidying up.

> > Maybe this should employ similar checks as __do_page_fault() (with
> > expand_stack() and VM_GROWSDOWN).
>
> You mean the code needs to detect endianness and to check if this is an
> attempt to expand the stack for both cases?

Nothing to do with endianness, just the relation between the address and
the vma->vm_start and whether the vma can be expanded down.

--
Catalin