ucount: use-after-free read in inc_ucount & dec_ucount

Dmitry Vyukov dvyukov at google.com
Sat Mar 4 12:01:09 UTC 2017


On Sat, Mar 4, 2017 at 12:50 PM, Dmitry Vyukov <dvyukov at google.com> wrote:
>> <n.borisov.lkml at gmail.com> wrote:
>>> [Addressing Dmitry Vyukov to ask for syzkaller clarification]
>>>
>>> On  3.03.2017 18:30, Eric W. Biederman wrote:
>>>> Nikolay Borisov <n.borisov.lkml at gmail.com> writes:
>>>>
>>>>> [Added containers ml, Eric Biederman and Jan Kara]. Please,
>>>>> next time don't add random people but take the time to see who touched
>>>>> the code.
>>>>
>>>> Comments below.
>>>>
>>>>> On  3.03.2017 14:16, JongHwan Kim wrote:
>>>>>> I've got the following report with syzkaller fuzzer
>>>>>>
>>>>>>
>>>>>> Syzkaller hit 'KASAN: use-after-free Read in dec_ucount' bug on commit .
>>>>>>
>>>>>> ==================================================================
>>>>>> BUG: KASAN: use-after-free in __read_once_size
>>>>>> include/linux/compiler.h:254 [inline] at addr ffff88006d399bc4
>>>>>> BUG: KASAN: use-after-free in atomic_read
>>>>>> arch/x86/include/asm/atomic.h:26 [inline] at addr ffff88006d399bc4
>>>>>> BUG: KASAN: use-after-free in atomic_dec_if_positive
>>>>>> include/linux/atomic.h:616 [inline] at addr ffff88006d399bc4
>>>>>> BUG: KASAN: use-after-free in dec_ucount+0x1e5/0x210 kernel/ucount.c:217
>>>>>> at addr ffff88006d399bc4
>>>>>> Read of size 4 by task syz-executor3/19713
>>>>>> CPU: 1 PID: 19713 Comm: syz-executor3 Not tainted 4.10.0+ #4
>>>>>> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS
>>>>>> Ubuntu-1.8.2-1ubuntu1 04/01/2014
>>>>>> Call Trace:
>>>>>>  __dump_stack lib/dump_stack.c:15 [inline]
>>>>>>  dump_stack+0x115/0x1cf lib/dump_stack.c:51
>>>>>>  kasan_object_err+0x1c/0x70 mm/kasan/report.c:162
>>>>>>  print_address_description mm/kasan/report.c:200 [inline]
>>>>>>  kasan_report_error mm/kasan/report.c:289 [inline]
>>>>>>  kasan_report.part.1+0x20e/0x4e0 mm/kasan/report.c:311
>>>>>>  kasan_report mm/kasan/report.c:331 [inline]
>>>>>>  __asan_report_load4_noabort+0x29/0x30 mm/kasan/report.c:331
>>>>>>  __read_once_size include/linux/compiler.h:254 [inline]
>>>>>>  atomic_read arch/x86/include/asm/atomic.h:26 [inline]
>>>>>>  atomic_dec_if_positive include/linux/atomic.h:616 [inline]
>>>>>>  dec_ucount+0x1e5/0x210 kernel/ucount.c:217
>>>>>>  dec_inotify_instances fs/notify/inotify/inotify.h:37 [inline]
>>>>>>  inotify_free_group_priv+0x6c/0x80 fs/notify/inotify/inotify_fsnotify.c:169
>>>>>>  fsnotify_final_destroy_group fs/notify/group.c:37 [inline]
>>>>>>  fsnotify_put_group+0x73/0xa0 fs/notify/group.c:110
>>>>>>  fsnotify_destroy_group+0xec/0x120 fs/notify/group.c:93
>>>>>>  inotify_release+0x37/0x50 fs/notify/inotify/inotify_user.c:280
>>>>>>  __fput+0x327/0x7e0 fs/file_table.c:208
>>>>>>  ____fput+0x15/0x20 fs/file_table.c:244
>>>>>>  task_work_run+0x18a/0x260 kernel/task_work.c:116
>>>>>>  exit_task_work include/linux/task_work.h:21 [inline]
>>>>>>  do_exit+0xa45/0x1b20 kernel/exit.c:873
>>>>>>  do_group_exit+0x149/0x400 kernel/exit.c:977
>>>>>>  get_signal+0x7d5/0x1810 kernel/signal.c:2313
>>>>>>  do_signal+0x94/0x1f30 arch/x86/kernel/signal.c:807
>>>>>>  exit_to_usermode_loop+0x162/0x1e0 arch/x86/entry/common.c:156
>>>>>>  prepare_exit_to_usermode arch/x86/entry/common.c:190 [inline]
>>>>>>  syscall_return_slowpath+0x2b6/0x310 arch/x86/entry/common.c:259
>>>>>>  entry_SYSCALL_64_fastpath+0xc0/0xc2
>>>>>
>>>>> So PID 19713 is exitting and as part of it it's freeing its file
>>>>> descriptors, one of which is apparently an inotify fd. And this has
>>>>> already been freed.
>>>>>
>>>>>
>>>>>> RIP: 0033:0x44fb79
>>>>>> RSP: 002b:00007ffd0f00f6d8 EFLAGS: 00000206 ORIG_RAX: 00000000000000ca
>>>>>> RAX: fffffffffffffdfc RBX: 0000000000708024 RCX: 000000000044fb79
>>>>>> RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000708024
>>>>>> RBP: 00000000000ae8e6 R08: 0000000000708000 R09: 000000160000000d
>>>>>> R10: 00007ffd0f00f710 R11: 0000000000000206 R12: 0000000000708000
>>>>>> R13: 0000000000708024 R14: 00000000000ae8a1 R15: 0000000000000016
>>>>>> Object at ffff88006d399b88, in cache kmalloc-96 size: 96
>>>>>> Allocated:
>>>>>> PID = 19691
>>>>>>  save_stack_trace+0x16/0x20 arch/x86/kernel/stacktrace.c:57
>>>>>>  save_stack+0x43/0xd0 mm/kasan/kasan.c:502
>>>>>>  set_track mm/kasan/kasan.c:514 [inline]
>>>>>>  kasan_kmalloc+0xad/0xe0 mm/kasan/kasan.c:605
>>>>>>  kmem_cache_alloc_trace+0xfb/0x280 mm/slub.c:2745
>>>>>>  kmalloc include/linux/slab.h:490 [inline]
>>>>>>  kzalloc include/linux/slab.h:663 [inline]
>>>>>>  get_ucounts kernel/ucount.c:140 [inline]
>>>>>>  inc_ucount+0x538/0xa70 kernel/ucount.c:195
>>>>>>  inotify_new_group+0x309/0x410 fs/notify/inotify/inotify_user.c:655
>>>>>>  SYSC_inotify_init1 fs/notify/inotify/inotify_user.c:682 [inline]
>>>>>>  SyS_inotify_init1 fs/notify/inotify/inotify_user.c:669 [inline]
>>>>>>  sys_inotify_init+0x17/0x80 fs/notify/inotify/inotify_user.c:696
>>>>>>  entry_SYSCALL_64_fastpath+0x1f/0xc2
>>>>>
>>>>> However, it has been actually allocated by a different process with pid
>>>>> 19691.
>>>>>
>>>>>
>>>>>> Freed:
>>>>>> PID = 19708
>>>>>>  save_stack_trace+0x16/0x20 arch/x86/kernel/stacktrace.c:57
>>>>>>  save_stack+0x43/0xd0 mm/kasan/kasan.c:502
>>>>>>  set_track mm/kasan/kasan.c:514 [inline]
>>>>>>  kasan_slab_free+0x73/0xc0 mm/kasan/kasan.c:578
>>>>>>  slab_free_hook mm/slub.c:1357 [inline]
>>>>>>  slab_free_freelist_hook mm/slub.c:1379 [inline]
>>>>>>  slab_free mm/slub.c:2961 [inline]
>>>>>>  kfree+0xe8/0x2c0 mm/slub.c:3882
>>>>>>  put_ucounts+0x1dd/0x270 kernel/ucount.c:172
>>>>>>  dec_ucount+0x172/0x210 kernel/ucount.c:220
>>>>>>  dec_inotify_instances fs/notify/inotify/inotify.h:37 [inline]
>>>>>>  inotify_free_group_priv+0x6c/0x80 fs/notify/inotify/inotify_fsnotify.c:169
>>>>>>  fsnotify_final_destroy_group fs/notify/group.c:37 [inline]
>>>>>>  fsnotify_put_group+0x73/0xa0 fs/notify/group.c:110
>>>>>>  fsnotify_destroy_group+0xec/0x120 fs/notify/group.c:93
>>>>>>  inotify_release+0x37/0x50 fs/notify/inotify/inotify_user.c:280
>>>>>>  __fput+0x327/0x7e0 fs/file_table.c:208
>>>>>>  ____fput+0x15/0x20 fs/file_table.c:244
>>>>>>  task_work_run+0x18a/0x260 kernel/task_work.c:116
>>>>>>  exit_task_work include/linux/task_work.h:21 [inline]
>>>>>>  do_exit+0xa45/0x1b20 kernel/exit.c:873
>>>>>>  do_group_exit+0x149/0x400 kernel/exit.c:977
>>>>>>  get_signal+0x7d5/0x1810 kernel/signal.c:2313
>>>>>>  do_signal+0x94/0x1f30 arch/x86/kernel/signal.c:807
>>>>>>  exit_to_usermode_loop+0x162/0x1e0 arch/x86/entry/common.c:156
>>>>>>  prepare_exit_to_usermode arch/x86/entry/common.c:190 [inline]
>>>>>>  syscall_return_slowpath+0x2b6/0x310 arch/x86/entry/common.c:259
>>>>>>  entry_SYSCALL_64_fastpath+0xc0/0xc2
>>>>>
>>>>> And yet we have a third process which freed it, PID 19708. So there is
>>>>> some dance happening with this fd, being allocated by one process,
>>>>> handed over to 2 more, which are freeing it. Is this a valid usage
>>>>> scenario of inotify descriptors?
>>>>
>>>> They are file descriptors so passing them around is valid.  That is
>>>> something unix domain sockets have allowed since the dawn of linux.
>>>>
>>>> The dance would need to be the fd being passed to the addtional
>>>> processes and then closed in the original before being closed
>>>> in the processes the fd was passed to.
>>>>
>>>> If those additional processes last longer than the original process this
>>>> is easy to achieve.
>>>>
>>>> My guess is that someone just taught syskallzer to pass file descriptors
>>>> around.  So this may be an old bug.  Either that or syskallzer hasn't
>>>> been looking at linux-next with KASAN enabled in the kernel.
>>>
>>> Dmitry, can you tell if syzkaller tests sending file descriptors across
>>> sockets? Since the calltraces here show multiple processes being
>>> involved in different operations on the exact same file descriptor.
>>>
>>> Also JongHwan, can you provide the full, compilable reproducer to try
>>> and track this issue down?
>>
>>
>> syzkaller can pass descriptors across sockets, but currently only
>> within a single multi-threaded process.
>>
>> Are you sure it's the same descriptor? It seems to me that it's struct
>> ucounts, which is shared via the global ucounts_hashtable, so no
>> sharing is required in user processes.
>>
>> Unless I am missing something, we want:
>>
>> @@ -154,7 +155,7 @@ static struct ucounts *get_ucounts(struct
>> user_namespace *ns, kuid_t uid)
>>                         ucounts = new;
>>                 }
>>         }
>> -       if (!atomic_add_unless(&ucounts->count, 1, INT_MAX))
>> +       if (!atomic_add_unless(&ucounts->count, 1, 0))
>>                 ucounts = NULL;
>>         spin_unlock_irq(&ucounts_lock);
>>         return ucounts;
>>
>> no?
>>
>> put_ucounts drops the last reference, then get_ucounts finds the
>> ucounts and successfully increments refcount as it's not INT_MAX (it's
>> 0) and starts using it, meanwhile put_ucounts proceeds to
>> unconditionally deleting the ucounts.
>
>
> It also seems that a concurrent put_ucounts can make get_ucounts fail
> _spuriously_, which does not look good.
> Don't we want something along the following lines?
>
> diff --git a/kernel/ucount.c b/kernel/ucount.c
> index 8a11fc0cb459..233c8e46acd5 100644
> --- a/kernel/ucount.c
> +++ b/kernel/ucount.c
> @@ -143,19 +143,18 @@ static struct ucounts *get_ucounts(struct
> user_namespace *ns, kuid_t uid)
>
>                 new->ns = ns;
>                 new->uid = uid;
> -               atomic_set(&new->count, 0);
> +               atomic_set(&new->count, 1);
>
>                 spin_lock_irq(&ucounts_lock);
>                 ucounts = find_ucounts(ns, uid, hashent);
>                 if (ucounts) {
> +                       atomic_inc(&ucounts->count);
>                         kfree(new);
>                 } else {
>                         hlist_add_head(&new->node, hashent);
>                         ucounts = new;
>                 }
>         }
> -       if (!atomic_add_unless(&ucounts->count, 1, INT_MAX))
> -               ucounts = NULL;
>         spin_unlock_irq(&ucounts_lock);
>         return ucounts;
>  }
> @@ -166,7 +165,10 @@ static void put_ucounts(struct ucounts *ucounts)
>
>         if (atomic_dec_and_test(&ucounts->count)) {
>                 spin_lock_irqsave(&ucounts_lock, flags);
> -               hlist_del_init(&ucounts->node);
> +               if (atomic_read(&ucounts->count) == 0)
> +                       hlist_del_init(&ucounts->node);
> +               else
> +                       ucounts = NULL;
>                 spin_unlock_irqrestore(&ucounts_lock, flags);
>
>                 kfree(ucounts);


/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\

This is broken per se. Need something more elaborate.


More information about the Containers mailing list