Re: gadgetfs broken since 7f7f25e8

[Alexander Holler <holler@ahsoftware.de>]

Am 02.03.2015 um 12:39 schrieb Alexander Holler:
> Am 02.03.2015 um 11:20 schrieb Al Viro:
>> On Mon, Mar 02, 2015 at 10:13:27AM +0100, Richard Weinberger wrote:
>>> On Mon, Mar 2, 2015 at 9:28 AM, Alexander Holler
>>> <holler@ahsoftware.de> wrote:
>>>> Hello.
>>>>
>>>> Commit 7f7f25e82d54870df24d415a7007fbd327da027b (introduced with
>>>> 3.16) broke
>>>> dynamic changing of file_operations->[read|write].
>>>>
>>>> At least gadgetfs is a victim.
>>>>
>>>> Feel free to ask me off-list for a patch as I don't want to end up in
>>>> annoying discussions on Linux kernel lists anymore.
>>>>
>>>> Alexander Holler
>>>
>>> CC'ing Al.
>>
>> I know.  FWIW, gadgetfs is one of the very few places that tried to
>> pull that
>> crap off and it had always been seriously racy.  I've posted a partial
>> analysis
>> about a month ago (<20150204190645.GJ29656@ZenIV.linux.org.uk>).
>>
>> If Alexander (or anybody else) has a patch that really fixes that thing,
>> I would certainly like to see it.  If not, I'll try to cook something,
>> but I'm not very familiar with that code.  I really hope that this patch
>> isn't "modify ->f_mode to match ->f_op change" - that's too racy.
>> We'll obviously need to fix the userland-visible breakage in that one,
>> but that's not the way to go...
>
> I exactly did what you've assumed, I've just fixed f_mode but not the
> already existing races which I haven't introduced. So I was right in not
> sending a patch as would have been blamed for not rewriting everything
> as so often.

Another quick solution would be to just add some dummy ops for 
read/write to those file_operations which are missing it which are only 
returning -EINVAL or some other error which might make sense.

But that still won't fix any existing race occuring while changing the 
the ops.