Re: A small window for a race condition in mm/rmap.c:page_lock_anon_vma_read

[Hugh Dickins]

On Thu, 22 Dec 2016, Peter Zijlstra wrote:
> On Wed, Dec 21, 2016 at 03:43:43PM +0100, Michal Hocko wrote:
> > anon_vma locking is clever^Wsubtle as hell. CC Peter...
> > 
> > On Tue 20-12-16 09:32:27, Dashi DS1 Cao wrote:
> > > I've collected four crash dumps with similar backtrace. 
> > > 
> > > PID: 247    TASK: ffff881fcfad8000  CPU: 14  COMMAND: "kswapd1"
> > >  #0 [ffff881fcfad7978] machine_kexec at ffffffff81051e9b
> > >  #1 [ffff881fcfad79d8] crash_kexec at ffffffff810f27e2
> > >  #2 [ffff881fcfad7aa8] oops_end at ffffffff8163f448
> > >  #3 [ffff881fcfad7ad0] die at ffffffff8101859b
> > >  #4 [ffff881fcfad7b00] do_general_protection at ffffffff8163ed3e
> > >  #5 [ffff881fcfad7b30] general_protection at ffffffff8163e5e8
> > >     [exception RIP: down_read_trylock+9]
> > >     RIP: ffffffff810aa9f9  RSP: ffff881fcfad7be0  RFLAGS: 00010286
> > >     RAX: 0000000000000000  RBX: ffff882b47ddadc0  RCX: 0000000000000000
> > >     RDX: 0000000000000000  RSI: 0000000000000000  RDI: 91550b2b32f5a3e8
> > 
> > rdi is obviously a mess - smells like a string. So either sombody has
> > overwritten root_anon_vma or this is really a use after free...
> 
> e8 - ???
> a3 - ???
> f5 - ???
> 32 - 2
> 2b - +
>  b - 
> 
> 55 - U
> 91 - ???
> 
> Not a string..
> 
> > >     RBP: ffff881fcfad7be0   R8: ffffea00ecc28860   R9: ffff883fcffeae28
> > >     R10: ffffffff81a691a0  R11: 0000000000000001  R12: ffff882b47ddadc1
> > >     R13: ffffea00ecc28840  R14: 91550b2b32f5a3e8  R15: ffffea00ecc28840
> > >     ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0000
> > >  #6 [ffff881fcfad7be8] page_lock_anon_vma_read at ffffffff811a3365
> > >  #7 [ffff881fcfad7c18] page_referenced at ffffffff811a35e7
> > >  #8 [ffff881fcfad7c90] shrink_active_list at ffffffff8117e8cc
> > >  #9 [ffff881fcfad7d48] balance_pgdat at ffffffff81180288
> > > #10 [ffff881fcfad7e20] kswapd at ffffffff81180813
> > > #11 [ffff881fcfad7ec8] kthread at ffffffff810a5b8f
> > > #12 [ffff881fcfad7f50] ret_from_fork at ffffffff81646a98
> > > 
> > > I suspect my customer hits into a small window of a race condition in mm/rmap.c: page_lock_anon_vma_read.
> > > struct anon_vma *page_lock_anon_vma_read(struct page *page)
> > > {
> > >         struct anon_vma *anon_vma = NULL;
> > >         struct anon_vma *root_anon_vma;
> > >         unsigned long anon_mapping;
> > > 
> > >         rcu_read_lock();
> > >         anon_mapping = (unsigned long)READ_ONCE(page->mapping);
> > >         if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
> > >                 goto out;
> > >         if (!page_mapped(page))
> > >                 goto out;
> > > 
> > >         anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
> > >         root_anon_vma = READ_ONCE(anon_vma->root);
> > 
> > Could you dump the anon_vma and struct page as well?
> > 
> > >         if (down_read_trylock(&root_anon_vma->rwsem)) {
> > >                 /*
> > >                  * If the page is still mapped, then this anon_vma is still
> > >                  * its anon_vma, and holding the mutex ensures that it will
> > >                  * not go away, see anon_vma_free().
> > >                  */
> > >                 if (!page_mapped(page)) {
> > >                         up_read(&root_anon_vma->rwsem);
> > >                         anon_vma = NULL;
> > >                 }
> > >                 goto out;
> > >         }
> > > ...
> > > }
> > > 
> > > Between the time the two "page_mapped(page)" are checked, the address
> > > (anon_mapping - PAGE_MAPPING_ANON) is unmapped! However it seems
> > > that anon_vma->root could still be read in but the value is wild. So
> > > the kernel crashes in down_read_trylock. But it's weird that all the
> > > "struct page" has its member "_mapcount" still with value 0, not -1,
> > > in the four crashes.
> 
> So the point is that while we hold rcu_read_lock() the actual memory
> backing the anon_vmas cannot be freed. It can be reused, but only for
> another anon_vma.
> 
> Now, anon_vma_alloc() sets ->root to self, while anon_vma_free() leaves
> ->root set to whatever. And any other ->root assignment is to a valid
> anon_vma.
> 
> Therefore, the same rules that ensure anon_vma stays valid, should also
> ensure anon_vma->root stays valid.
> 
> Now, one thing that might go wobbly is that ->root assignments are not
> done using WRITE_ONCE(), this means a naughty compiler can miscompile
> those stores and introduce store-tearing, if our READ_ONCE() would
> observe such a tear, we'd be up some creek without a paddle.

We would indeed.  And this being the season of goodwill, I'm biting
my tongue not to say what I think of the prospect of store tearing.
But that zeroed anon_vma implies tearing not the problem here anyway.

> 
> Now, its been a long time since I looked at any of this code, and I see
> that Hugh has fixed at least two wobblies in my original code.

Nothing much, and this (admittedly subtle) technique has been working
well for years, so I'm sceptical about "a small window for a race
condition".

But Dashi's right to point out that the struct page has _mapcount 0
(not -1 for logical 0) in these cases: it looks as if something is
freeing (or corrupting) the anon_vma despite it still having pages
mapped, or something is misaccounting (or corrupting) the _mapcount.

But I've no idea what, and we have not heard such reports elsewhere.
We don't even know what kernel this is - something special, perhaps?

Hugh