Re: debugfs question...

[Nicolai Stange <nicstange@gmail.com>]

Hi again,

bad news: my previous analysis was completely wrong, c.f. below.
Good news (from my point of view): debugfs is correct, no fix needed for
it.

Apologies for the confusion...


Nicolai Stange <nicstange@gmail.com> writes:

> Greg KH <greg@kroah.com> writes:
>
>> On Mon, Oct 31, 2016 at 02:32:56PM -0400, Mike Marshall wrote:
>>
>>> But... really bad things happen if someone unloads the Orangefs
>>> module after my test program does the open and before the read
>>> starts. So I picked another debugfs-using-filesystem (f2fs) and
>>> pointed my tester-program at /sys/kernel/debug/f2fs/status, and
>>> the same bad thing happens there.
>
> [...]
>
>>> [ 1240.144316] Call Trace:
>>> [ 1240.144450]  [<ffffffff8122907f>] __fput+0xdf/0x1d0
>>> [ 1240.144704]  [<ffffffff812291ae>] ____fput+0xe/0x10
>>> [ 1240.144962]  [<ffffffff810b97de>] task_work_run+0x8e/0xc0
>>> [ 1240.145243]  [<ffffffff8109b98e>] do_exit+0x2ae/0xae0
>
>
> Thank you very much for this detailed report!
>
> At least for the .../f2fs/status file, your splat at fput() can be
> readily explained with the full proxy's releaser not being protected
> against file removals in any way.
>
> Partly this is on purpose, c.f. the comment in full_proxy_release().
>
> However, I should have at least tried to acquire a reference to the
> owning module before accessing some static struct file_operations or
> even calling some ->release() within it. Meh.

This is what I got wrong: debugfs does acquire the needed references
correctly (details below). For the case of f2fs' "status" file, the
file_operations ->owner is simply not set as it should have been,
i.e. to THIS_MODULE.


Details on debugfs' reference acquisition:
The open proxy, full_proxy_open(), gets a reference to the "real"
file_operations, hence to its module. (Only in its error path it
releases it again).

full_proxy_release() is in charge of dropping that reference again, but
only *after* it has attempted to call the "real" ->release().

So, as long as a file has been (successfully) opened, a reference to the
original file_operation's ->owner is owned, preventing it from getting
unloaded.


Can you confirm that you didn't set ->owner in your Orangefs related
tests, too?


Thanks,

Nicolai