Re: [RFC PATCH 1/3] debugfs: fix automount inode i_nlink references

[Roman Peniaev]

On Mon, Feb 8, 2016 at 5:35 PM, Greg Kroah-Hartman
<gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> On Mon, Feb 08, 2016 at 11:28:52AM +0100, Roman Peniaev wrote:
>> On Mon, Feb 8, 2016 at 7:38 AM, Greg Kroah-Hartman
>> <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
>> > On Thu, Dec 10, 2015 at 01:47:12PM +0100, Roman Pen wrote:
>> >> Directory inodes should start off with i_nlink == 2 (for "." entry).
>> >> Of course the same rule should be applied to automount dentries for
>> >> child and parent inodes as well.
>> >>
>> >> Also now automount dentry does fsnotify_mkdir.
>> >>
>> >> Without this patch kernel complains when sees i_nlink == 0:
>> >
>> > How can the kernel see this?  What did you do to trigger this?
>>
>> Yes, sorry, I had to be more precise on this.
>> That happens on unlinking of automount dentry.
>>
>> Easily can be reproduced:
>>
>>     autom = debugfs_create_automount("automount", parentd, vfsmount_cb, data);
>>     BUG_ON(IS_ERR_OR_NULL(autom));
>>     debugfs_remove(autom);
>>
>> You will immediately see one warning on attempt to drop_nlink() (which is zero)
>> for automount dentry.
>
> Why don't we see this "in the wild" today with the one user of this
> function?

I see only one current user: tracefs.
tracefs does not remove the automount root dentry.
That's the reason.

>
>> The second warning happens when you unlink 'parentd', because
>> debugfs_create_automount() did not increase the nlink for parent
>> inode.
>>
>> Do I need to resend this patch with more precise description?
>
> Yes, please fix up and resend as a stand-alone patch, as it is
> independant of your other proposal.
>
> And take off the "RFC" marking, I can never apply a patch with that type
> of marking as you obviously don't think it is good enough to be merged,
> so why would I?  :)

I've resent a patch.
Please, take a look.

--
Roman