Re: [PATCH bpf 1/3] bpf: Pin the start cgroup in cgroup_iter_seq_init()

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi,

On 11/8/2022 3:03 PM, Yonghong Song wrote:
>
>
> On 11/7/22 8:08 PM, Hou Tao wrote:
>> Hi,
>>
>> On 11/8/2022 10:11 AM, Hao Luo wrote:
>>> On Mon, Nov 7, 2022 at 1:59 PM Yonghong Song <yhs@xxxxxxxx> wrote:
>>>>
>>>>
>>>> On 11/6/22 11:42 PM, Hou Tao wrote:
>>>>> From: Hou Tao <houtao1@xxxxxxxxxx>
>>>>>
>>>>> bpf_iter_attach_cgroup() has already acquired an extra reference for the
>>>>> start cgroup, but the reference will be released if the iterator link fd
>>>>> is closed after the creation of iterator fd, and it may lead to
>>>>> User-After-Free when reading the iterator fd.
>>>>>
>>>>> So fixing it by acquiring another reference for the start cgroup.
>>>>>
>>>>> Fixes: d4ccaf58a847 ("bpf: Introduce cgroup iter")
>>>>> Signed-off-by: Hou Tao <houtao1@xxxxxxxxxx>
>>>> Acked-by: Yonghong Song <yhs@xxxxxx>
>>> There is an alternative: does it make sense to have the iterator hold
>>> a ref of the link? When the link is closed, my assumption is that the
>>> program is already detached from the cgroup. After that, it makes no
>>> sense to still allow iterating the cgroup. IIUC, holding a ref to the
>>> link in the iterator also fixes for other types of objects.
>> Also considered the alternative solution when fixing the similar problem in bpf
>> map element iterator [0]. The problem is not all of bpf iterators need the
>> pinning (e.g., bpf map iterator). Because bpf prog is also pinned by iterator fd
>> in iter_open(), so closing the fd of iterator link doesn't release the bpf
>> program.
>>
>> [0]: https://lore.kernel.org/bpf/20220810080538.1845898-2-houtao@xxxxxxxxxxxxxxx/
>
> Okay, let us do the solution to hold a reference to the link for the iterator.
> For cgroup_iter, that means, both prog and cgroup will be present so we should
> be okay then.
The reason I did not use the solution is that it will create unnecessary
dependency between iterator fd and iterator link and many bpf iterators also
don't need that. If we use the solution, should I revert the fixes to bpf map
iterator done before or keep it as-is ?
>
>>>
>>> Hao
>>




[Index of Archives]     [Linux Samsung SoC]     [Linux Rockchip SoC]     [Linux Actions SoC]     [Linux for Synopsys ARC Processors]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]


  Powered by Linux