Re: [PATCH] kernfs: Fix UAF in PSI polling when open file is released

[Chen Ridong <chenridong@huaweicloud.com>]

On 2025/8/15 22:42, Michal Koutný wrote:
> On Fri, Aug 15, 2025 at 08:11:39AM +0200, Greg KH <gregkh@linuxfoundation.org> wrote:
>>> diff --git a/fs/kernfs/file.c b/fs/kernfs/file.c
>>> index a6c692cac616..d5d01f0b9392 100644
>>> --- a/fs/kernfs/file.c
>>> +++ b/fs/kernfs/file.c
>>> @@ -852,7 +852,7 @@ static __poll_t kernfs_fop_poll(struct file *filp, poll_table *wait)
>>>  	struct kernfs_node *kn = kernfs_dentry_node(filp->f_path.dentry);
>>>  	__poll_t ret;
>>>  
>>> -	if (!kernfs_get_active(kn))
>>> +	if (of->released || !kernfs_get_active(kn))
>>
>> I can see why the cgroup change is needed,
> 
> I don't see it that much. of->priv isn't checked in cgroup code anywhere
> so it isn't helpful zeroing. As Ridong writes it may trade UaF for NULL
> pointer deref :-/ (Additionally, same zeroing would be needed in error
> path in cgroup_file_open().)
> 

Thank you, Michal,

I believe assigning NULL to of->priv should be harmless. This change would make the bug more
observable in practice. Without this explicit NULL assignment, the use-after-free (UAF) issue might
remain hidden in some cases, particularly when KASAN is disabled.

> I _think_ the place to cleanup would be in
> @@ -3978,6 +3978,8 @@ static ssize_t cgroup_pressure_write(struct kernfs_open_file *of,
>                 psi->enabled = enable;
>                 if (enable)
>                         psi_cgroup_restart(psi);
> +               else
> +                       psi_trigger_destroy(???);
>         }
> 

Could you please provide more details about this modification? Do you mean we need to consider
additional cleanup work when disabling cgroup.pressure? The psi_trigger_destroy is invoked as
follows:

cgroup_file_show
  kernfs_drain
    kernfs_drain_open_files
      kernfs_release_file
        cgroup_file_release
          cft->release(of);
            cgroup_pressure_release
              psi_trigger_destroy

>         cgroup_kn_unlock(of->kn);
> 
> The issue is that cgroup_pressure_write doesn't know all possible
> triggers to be cancelled. (The fix with of->released would only
> sanitize effect but not the cause IMO.)
> 
> HTH,
> Michal

-- 
Best regards,
Ridong