io_close()

io_close()

[Pavel Begunkov]

[-- Attachment #1.1: Type: text/plain, Size: 297 bytes --]

Hi,

I noticed, that io_close() is broken for some use cases, and was thinking about
the best way to fix it. Is fput(req->close.put_file) really need to be done in
wq? It seems, fput_many() implementation just calls schedule_delayed_work(), so
it's already delayed.

-- 
Pavel Begunkov


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: io_close()

[Jens Axboe]

On 2/7/20 1:52 PM, Pavel Begunkov wrote:
> Hi,
> 
> I noticed, that io_close() is broken for some use cases, and was thinking about
> the best way to fix it. Is fput(req->close.put_file) really need to be done in
> wq? It seems, fput_many() implementation just calls schedule_delayed_work(), so
> it's already delayed.

It's not the fput(), it's the f_op->flush().

-- 
Jens Axboe

Re: io_close()

[Pavel Begunkov]

[-- Attachment #1.1: Type: text/plain, Size: 472 bytes --]

On 08/02/2020 00:14, Jens Axboe wrote:
> On 2/7/20 1:52 PM, Pavel Begunkov wrote:
>> Hi,
>>
>> I noticed, that io_close() is broken for some use cases, and was thinking about
>> the best way to fix it. Is fput(req->close.put_file) really need to be done in
>> wq? It seems, fput_many() implementation just calls schedule_delayed_work(), so
>> it's already delayed.
> 
> It's not the fput(), it's the f_op->flush().
> 

Got it, thanks

-- 
Pavel Begunkov


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: io_close()

[Pavel Begunkov]

[-- Attachment #1.1: Type: text/plain, Size: 711 bytes --]

On 08/02/2020 00:14, Jens Axboe wrote:
> On 2/7/20 1:52 PM, Pavel Begunkov wrote:
>> Hi,
>>
>> I noticed, that io_close() is broken for some use cases, and was thinking about
>> the best way to fix it. Is fput(req->close.put_file) really need to be done in
>> wq? It seems, fput_many() implementation just calls schedule_delayed_work(), so
>> it's already delayed.
> 
> It's not the fput(), it's the f_op->flush().

What confuses me, is that in case of ->flush, it doesn't return -EAGAIN, but
io_queue_async_work() itself, so nobody will set ->work.files. But that means
io_close_finish() won't do filp_close() as well.

Did I missed somewhere called io_grab_files()?


-- 
Pavel Begunkov


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]