Re: Multiple potential races on vma->vm_flags

[Oleg Nesterov]

On 09/24, Andrey Ryabinin wrote:
>
> 2015-09-24 20:26 GMT+03:00 Oleg Nesterov <oleg@xxxxxxxxxx>:
> > On 09/24, Sasha Levin wrote:
> >>
> >> void unmap_vmas(struct mmu_gather *tlb,
> >>                 struct vm_area_struct *vma, unsigned long start_addr,
> >>                 unsigned long end_addr)
> >> {
> >>         struct mm_struct *mm = vma->vm_mm;
> >>
> >>         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
> >>         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
> >>                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL); <--- this
> >>         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
> >> }
> >
> > And I do not see any dereference at this line,
> >
>
> I noticed, that addr2line sometimes doesn't work reliably on
> compiler-instrumented code.
> I've seen couple times that it points to the next line of code.

Yes, I know that we can't trust it. That is why I think (at least in
this particular case) function+offset would be more helpful. And we
need more asm probably.

> >> >>    0:   08 80 3c 02 00 0f       or     %al,0xf00023c(%rax)
> >> >>    6:   85 22                   test   %esp,(%rdx)
> >> >>    8:   01 00                   add    %eax,(%rax)
> >> >>    a:   00 48 8b                add    %cl,-0x75(%rax)
> >> >>    d:   43                      rex.XB
> >> >>    e:   40                      rex
> >> >>    f:   48 8d b8 c8 04 00 00    lea    0x4c8(%rax),%rdi
> >> >>   16:   48 89 45 d0             mov    %rax,-0x30(%rbp)
> >> >>   1a:   48 b8 00 00 00 00 00    movabs $0xdffffc0000000000,%rax
> >> >>   21:   fc ff df
> >> >>   24:   48 89 fa                mov    %rdi,%rdx
> >> >>   27:   48 c1 ea 03             shr    $0x3,%rdx
> >> >>   2b:*  80 3c 02 00             cmpb   $0x0,(%rdx,%rax,1)               <-- trapping instruction
> >> >>   2f:   0f 85 ee 00 00 00       jne    0x123
> >> >>   35:   48 8b 45 d0             mov    -0x30(%rbp),%rax
> >> >>   39:   48 83 b8 c8 04 00 00    cmpq   $0x0,0x4c8(%rax)
> >> >>   40:   00
> >> >
> >> > And I do not see anything similar in "objdump -d". So could you at least
> >> > show mm/memory.c:1337 in your tree?
> >> >
> >> > Hmm. movabs $0xdffffc0000000000,%rax above looks suspicious, this looks
> >> > like kasan_mem_to_shadow(). So perhaps this code was generated by kasan?
> >> > (I can't check, my gcc is very old). Or what?
> >>
> >> This is indeed kasan code. 0xdffffc0000000000 is the shadow base, and you see
> >> kasan trying to access shadow base + (ptr >> 3), which is why we get GFP.
> >
> > and thus this asm can't help, right?
> >
>
> I think it can.
>
> > So how can we figure out where exactly the kernel hits NULL ? And what
> > exactly it tries to dereference?
>
> So we tried to dereference 0x4c8.  That 0x4c8 is probably offset in some struct.
> The only big struct here is mm_struct.
> So I think that we tried to derefernce null mm, and this asm:
>          > cmpq   $0x0,0x4c8(%rax)
>
> is likely from inlined mm_has_notifiers():
>     static inline int mm_has_notifiers(struct mm_struct *mm)
>     {
>              return unlikely(mm->mmu_notifier_mm);
>     }

Looks reasonable... Thanks.

I was going to say that this is impossible because the caller should have
crashed if ->mm == NULL. But unmap_vmas() uses mm = vma->vm_mm, so it looks
like this vma or mm->mmap was corrupted...

Oleg.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/