Re: KASAN: use-after-free Write in detach_if_pending

From: Dmitry Vyukov
Date: Mon Oct 30 2017 - 11:48:57 EST


On Sun, Oct 29, 2017 at 2:01 PM, Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
> On Sun, 2017-10-29 at 13:45 +0100, Thomas Gleixner wrote:
>> On Fri, 27 Oct 2017, syzbot wrote:
>>
>> Cc'ed network folks.
>>
>> > syzkaller hit the following crash on e7989f973ae1b90ec7c0b671c81f7f553affccbe
>> > git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/master
>> > compiler: gcc (GCC) 7.1.1 20170620
>> > .config is attached
>> > Raw console output is attached.
>> > C reproducer is attached
>> > syzkaller reproducer is attached. See https://goo.gl/kgGztJ
>> > for information about syzkaller reproducers
>> >
>> >
>> > BUG: KASAN: use-after-free in __write_once_size include/linux/compiler.h:305
>> > [inline]
>> > BUG: KASAN: use-after-free in __hlist_del include/linux/list.h:648 [inline]
>> > BUG: KASAN: use-after-free in detach_timer kernel/time/timer.c:791 [inline]
>> > BUG: KASAN: use-after-free in detach_if_pending+0x557/0x610
>> > kernel/time/timer.c:808
>> > Write of size 8 at addr ffff8801d3bab780 by task syzkaller900516/2986
>>
>> That's just the point where this gets detected.
>>
>> > CPU: 1 PID: 2986 Comm: syzkaller900516 Not tainted 4.13.0+ #82
>>
>> > __hlist_del include/linux/list.h:648 [inline]
>> > detach_timer kernel/time/timer.c:791 [inline]
>> > detach_if_pending+0x557/0x610 kernel/time/timer.c:808
>> > try_to_del_timer_sync+0xa2/0x120 kernel/time/timer.c:1182
>> > del_timer_sync+0x18a/0x240 kernel/time/timer.c:1247
>> > tun_flow_uninit drivers/net/tun.c:1104 [inline]
>> > tun_free_netdev+0x105/0x1b0 drivers/net/tun.c:1776
>>
>> ^^^^^^^^^^^^ This shouldn't be called I think
>>
>> > netdev_run_todo+0x870/0xca0 net/core/dev.c:7864
>> > rtnl_unlock+0xe/0x10 net/core/rtnetlink.c:106
>> > tun_detach drivers/net/tun.c:588 [inline]
>> > tun_chr_close+0x49/0x60 drivers/net/tun.c:2609
>> > __fput+0x333/0x7f0 fs/file_table.c:210
>> > ____fput+0x15/0x20 fs/file_table.c:246
>> > task_work_run+0x199/0x270 kernel/task_work.c:112
>> > exit_task_work include/linux/task_work.h:21 [inline]
>> > do_exit+0xa52/0x1b40 kernel/exit.c:865
>>
>> Here is the allocation path
>>
>> > alloc_netdev_mqs+0x16e/0xed0 net/core/dev.c:8018
>> > tun_set_iff drivers/net/tun.c:2022 [inline]
>> > __tun_chr_ioctl+0x12be/0x3d20 drivers/net/tun.c:2276
>> > tun_chr_ioctl+0x2a/0x40 drivers/net/tun.c:2521
>> > vfs_ioctl fs/ioctl.c:45 [inline]
>> > do_vfs_ioctl+0x1b1/0x1530 fs/ioctl.c:685
>> > SYSC_ioctl fs/ioctl.c:700 [inline]
>> > SyS_ioctl+0x8f/0xc0 fs/ioctl.c:691
>> > entry_SYSCALL_64_fastpath+0x1f/0xbe
>>
>>
>> And this is free.
>>
>> > netdev_freemem net/core/dev.c:7970 [inline]
>> > free_netdev+0x2cf/0x360 net/core/dev.c:8132
>> > tun_set_iff drivers/net/tun.c:2105 [inline]
>>
>> err_free_flow:
>> tun_flow_uninit(tun); <--------
>>
>> > __tun_chr_ioctl+0x2cf6/0x3d20 drivers/net/tun.c:2276
>> > tun_chr_ioctl+0x2a/0x40 drivers/net/tun.c:2521
>> > vfs_ioctl fs/ioctl.c:45 [inline]
>> > do_vfs_ioctl+0x1b1/0x1530 fs/ioctl.c:685
>> > SYSC_ioctl fs/ioctl.c:700 [inline]
>> > SyS_ioctl+0x8f/0xc0 fs/ioctl.c:691
>> > entry_SYSCALL_64_fastpath+0x1f/0xbe
>>
>> So it's the TUNSETIFF ioctl which first allocates and then frees in the
>> errorpath of tun_set_iff.
>>
>> But for some reason this sticks and the exit of that task does it again,
>> which triggers KASAN in the innocent timer code.
>
> Pretty old story, already fixed in David Miller trees.
>
> net-next tree :
>
> $ git log --oneline e7989f973ae1b90ec7c0b671c81.. -- drivers/net/tun.c
> f8ddadc4db6c7b7029b6d0e0d9af24f74ad27ca2 Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
> ee74d9967b829232723939cb7c9b100b29f6ec98 tun: do not arm flow_gc_timer in tun_flow_init()
> 81d98fa4df3d1683b3ef21e8a7a0ccac7874f0de tun: avoid extra timer schedule in tun_flow_cleanup()
> 7dbfb4ef77db5666f0f3a425e7db93ca30ff4285 tun: do not block BH again in tun_flow_cleanup()
> aec72f3392b1d598a979e89c4fdb131965ae0ab3 net-tun: fix panics at dismantle time
> 010f245b9dd734adda6386c494a4ace953ea8dc4 tun: relax check on eth_get_headlen() return value
> 0ad646c81b2182f7fa67ec0c8c825e0ee165696d tun: call dev_get_valid_name() before register_netdevice()
> 53954cf8c5d205624167a2bfd117cc0c1a5f3c6d Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
> 2580c4c17aee3ad58e9751012bad278dd074ccae tun: bail out from tun_get_user() if the skb is empty
> de8f3a83b0a0fddb2cf56e7a718127e9619ea3da bpf: add meta pointer for direct access
> 9484dc74fcf0750cd6726c9aa27edf97223916a8 tun: delete original tun_get() and rename __tun_get() to tun_get()
> 90e33d45940793def6f773b2d528e9f3c84ffdc7 tun: enable napi_gro_frags() for TUN/TAP driver
> 943170998b200190f99d3fe7e771437e2c51f319 tun: enable NAPI for TUN/TAP driver
>
> net tree :
>
> $ git log --oneline e7989f973ae1b90ec7c0b671c81.. -- drivers/net/tun.c
> 63b9ab65bd76e5de6479bb14b4014b64aa1a317a tuntap: properly align skb->head before building skb
> 5c25f65fd1e42685f7ccd80e0621829c105785d9 tun: allow positive return values on dev_get_valid_name() call
> 0ad646c81b2182f7fa67ec0c8c825e0ee165696d tun: call dev_get_valid_name() before register_netdevice()
> 2580c4c17aee3ad58e9751012bad278dd074ccae tun: bail out from tun_get_user() if the skb is empty
>
> Pick the fixes, they are at least 2 patches that addressed the issue.

Let's try the last one in net-next that touches timers:

#syz fix: tun: do not arm flow_gc_timer in tun_flow_init()