Re: Inquiry about the removal of flag O_NONBLOCK on /dev/random

["Guozihua (Scott)" <guozihua@huawei.com>]

On 2022/9/19 18:40, Jason A. Donenfeld wrote:
> Hi,
> 
> On Mon, Sep 19, 2022 at 11:27 AM Guozihua (Scott) <guozihua@huawei.com> wrote:
>>
>> On 2022/9/8 17:51, Jason A. Donenfeld wrote:
>>> Hi,
>>>
>>> On Thu, Sep 08, 2022 at 11:31:31AM +0800, Guozihua (Scott) wrote:
>>>> For example:
>>>>
>>>>
>>>> --
>>>> Best
>>>> GUO Zihua
>>>>
>>>> --
>>>> Best
>>>> GUO Zihua
>>>
>>> Looks like you forgot to paste the example...
>>>
>>>> Thank you for the timely respond and your patient. And sorry for the
>>>> confusion.
>>>>
>>>> First of all, what we think is that this change (removing O_NONBLOCK) is
>>>> reasonable. However, this do cause issue during the test on one of our
>>>> product which uses O_NONBLOCK flag the way I presented earlier in the
>>>> Linux 4.4 era. Thus our colleague suggests that returning -EINVAL when
>>>> this flag is received would be a good way to indicate this change.
>>>
>>> No, I don't think it's wise to introduce yet *new* behavior (your
>>> proposed -EINVAL). That would just exacerbate the (mostly) invisible
>>> breakage from the 5.6-era change.
>>>
>>> The question now before us is whether to bring back the behavior that
>>> was there pre-5.6, or to keep the behavior that has existed since 5.6.
>>> Accidental regressions like this (I assume it was accidental, at least)
>>> that are unnoticed for so long tend to ossify and become the new
>>> expected behavior. It's been around 2.5 years since 5.6, and this is the
>>> first report of breakage. But the fact that it does break things for you
>>> *is* still significant.
>>>
>>> If this was just something you noticed during idle curiosity but doesn't
>>> have a real impact on anything, then I'm inclined to think we shouldn't
>>> go changing the behavior /again/ after 2.5 years. But it sounds like
>>> actually you have a real user space in a product that stopped working
>>> when you tried to upgrade the kernel from 4.4 to one >5.6. If this is
>>> the case, then this sounds truly like a userspace-breaking regression,
>>> which we should fix by restoring the old behavior. Can you confirm this
>>> is the case? And in the meantime, I'll prepare a patch for restoring
>>> that old behavior.
>>>
>>> Jason
>>> .
>>
>> Hi Jason
>>
>> Thank for your patience.
>>
>> To answer your question, yes, we do have a userspace program reading
>> /dev/random during early boot which relies on O_NONBLOCK. And this
>> change do breaks it. The userspace program comes from 4.4 era, and as
>> 4.4 is going EOL, we are switching to 5.10 and the breakage is reported.
>>
>> It would be great if the kernel is able to restore this flag for
>> backward compatibility.
> 
> Alright then. Sounds like a clear case of userspace being broken. I'll
> include https://git.zx2c4.com/linux-rng/commit/?id=b931eaf6ef5cef474a1171542a872a5e270e3491
> or similar in my pull for 6.1, if that's okay with you. For 6.0, we're
> already at rc6, so maybe better to let this one stew for a bit longer,
> given the change, unless you feel strongly about having it earlier, I
> guess.
> 
> Jason
> .

Hi Jason

That's OK with us. Thanks.

-- 
Best
GUO Zihua