[EXT] Re: [PATCH 2/2] block: remove the per-bio/request write hint.

[Jens Axboe]

On 3/10/22 4:34 AM, Luca Porzio (lporzio) wrote:
>> -----Original Message-----
>> From: Manjong Lee <mj0123.lee at samsung.com>
>> Sent: Wednesday, March 9, 2022 2:31 PM
>> To: david at fromorbit.com
>> Cc: axboe at kernel.dk; hch at lst.de; kbusch at kernel.org; linux-
>> block at vger.kernel.org; linux-fsdevel at vger.kernel.org; linux-
>> nvme at lists.infradead.org; linux-raid at vger.kernel.org; sagi at grimberg.me;
>> song at kernel.org; seunghwan.hyun at samsung.com;
>> sookwan7.kim at samsung.com; nanich.lee at samsung.com;
>> woosung2.lee at samsung.com; yt0928.kim at samsung.com;
>> junho89.kim at samsung.com; jisoo2146.oh at samsung.com
>> Subject: [EXT] Re: [PATCH 2/2] block: remove the per-bio/request write hint.
>>
>> CAUTION: EXTERNAL EMAIL. Do not click links or open attachments unless
>> you recognize the sender and were expecting this message.
>>
>>
>>> On Sun, ddMar 06, 2022 at 11:06:12AM -0700, Jens Axboe wrote:
>>>> On 3/6/22 11:01 AM, Christoph Hellwig wrote:
>>>>> On Sun, Mar 06, 2022 at 10:11:46AM -0700, Jens Axboe wrote:
>>>>>> Yes, I think we should kill it. If we retain the inode hint, the
>>>>>> f2fs doesn't need a any changes. And it should be safe to make the
>>>>>> per-file fcntl hints return EINVAL, which they would on older kernels
>> anyway.
>>>>>> Untested, but something like the below.
>>>>>
>>>>> I've sent this off to the testing farm this morning, but EINVAL
>>>>> might be even better:
>>>>>
>>>>> https://urldefense.com/v3/__http://git.infradead.org/users/hch/bloc
>>>>> k.git/shortlog/refs/heads/more-hint-
>> removal__;!!KZTdOCjhgt4hgw!qsgy
>>>>>
>> oejchUYPeorpCL0Ov3jPGvXpXgxa7hpSCViD7XQy7uJDMDLo3U8v_bmoUtg$
>>>
>>> Yup, I like that.
>>>
>>>> I do think EINVAL is better, as it just tells the app it's not
>>>> available like we would've done before. With just doing zeroes, that
>>>> might break applications that set-and-verify. Of course there's also
>>>> the risk of that since we retain inode hints (so they work), but fail file
>> hints.
>>>> That's a lesser risk though, and we only know of the inode hints
>>>> being used.
>>>
>>> Agreed, I think EINVAL would be better here - jsut make it behave like
>>> it would on a kernel that never supported this functionality in the
>>> first place. Seems simpler to me for user applications if we do that.
>>>
>>> Cheers,
>>>
>>> Dave.
>>> --
>>> Dave Chinner
>>> david at fromorbit.com
>>>
>>
>> Currently, UFS device also supports hot/cold data separation and uses
>> existing write_hint code.
>>
>> In other words, the function is also being used in storage other than NVMe,
>> and if it is removed, it is thought that there will be an operation problem.
>>
>> If the code is removed, I am worried about how other devices that use the
>> function.
>>
>> Is there a good alternative?
> 
> Hi all,
> 
> I work for Micron UFS team. I confirm and support Manjong message above.
> There are UFS customers using custom write_hint in Android and due to the 
> "upstream first" policy from Google, if you remove write_hints in block device,
> The Android ecosystem will suffer this lack.
> 
> Can we revert back this decision? Or think of an alternative solution which 
> may work?

You do both realize that this is just the file specific hint? Inode
based hints will still work fine for UFS.

-- 
Jens Axboe