Re: [PATCH] f2fs: f2fs supports uncached buffered I/O read

[Jens Axboe <axboe@kernel.dk>]

On 7/30/25 7:58 PM, hanqi wrote:
> 
> ? 2025/7/30 23:20, Jens Axboe ??:
>> On 7/28/25 2:28 AM, hanqi wrote:
>>> ? 2025/7/28 16:07, Chao Yu ??:
>>>> On 7/28/25 16:03, hanqi wrote:
>>>>> ? 2025/7/28 15:38, Chao Yu ??:
>>>>>
>>>>>> On 7/25/25 15:53, Qi Han wrote:
>>>>>>> Jens has already completed the development of uncached buffered I/O
>>>>>>> in git [1], and in f2fs, uncached buffered I/O read can be enabled
>>>>>>> simply by setting the FOP_DONTCACHE flag in f2fs_file_operations.
>>>>>> IIUC, we may suffer lock issue when we call pwritev(.. ,RWF_DONTCACHE)?
>>>>>> as Jen mentioned in below path, right?
>>>>>>
>>>>>> soft-irq
>>>>>> - folio_end_writeback()
>>>>>>     - filemap_end_dropbehind_write()
>>>>>>      - filemap_end_dropbehind()
>>>>>>       - folio_unmap_invalidate()
>>>>>>        - lock i_lock
>>>>>>
>>>>>> Thanks,
>>>>> That's how I understand it.
>>>> So I guess we need to wait for the support RWF_DONTCACHE on write path, unless
>>>> you can walk around for write path in this patch.
>>>>
>>>> Thanks,
>>> I think the read and write paths can be submitted separately.
>>> Currently, uncached buffered I/O write requires setting the
>>> FGP_DONTCACHE flag when the filesystem allocates a folio. In
>>> f2fs, this is done in the following path:
>>>
>>> - write_begin
>>>   - f2fs_write_begin
>>>    - __filemap_get_folio
>>>    As I understand it, if we don't set the FGP_DONTCACHE flag here, this
>>> issue shouldn't occur.
>> It won't cause an issue, but it also won't work in the sense that the
>> intent is that if the file system doesn't support DONTCACHE, it would
>> get errored at submission time. Your approach would just ignore the flag
>> for writes, rather than return -EOPNOTSUPP as would be expected.
>>
>> You could potentially make it work just on the read side by having the
>> f2fs write submit side check DONTCACHE on the write side and error them
>> out.
> 
> Hi Jens,
> 
> Thank you for your suggestions. I am currently working on modifying
> F2FS to handle the dontcache unmap operation in a workqueue. I expect
> to submit the patch soon, after which F2FS should also support uncached
> buffer I/O writes.

Sounds good, that's the right approach. Userspace needs to be able to
rely on the fact that if RWF_DONTCACHE io is submitted without error
that the target does the right thing as well, barring bugs of course.

-- 
Jens Axboe