__io_file_supports_nowait for regular files

__io_file_supports_nowait for regular files

[Christoph Hellwig]

Hi Jens,

is there any good reason __io_file_supports_nowait checks the
blk_queue_nowait flag for regular files?  The FMODE_NOWAIT is set
for regular files that support nowait I/O, and should be all that
is needed.  Even with a block device that does not nonblocking
I/O some thing like reading from the page cache can be done
non-blocking.

Re: __io_file_supports_nowait for regular files

[Jens Axboe]

On 7/21/22 9:37 AM, Christoph Hellwig wrote:
> Hi Jens,
> 
> is there any good reason __io_file_supports_nowait checks the
> blk_queue_nowait flag for regular files?  The FMODE_NOWAIT is set
> for regular files that support nowait I/O, and should be all that
> is needed.  Even with a block device that does not nonblocking
> I/O some thing like reading from the page cache can be done
> non-blocking.

Nope, we can probably kill that check then. Want to send a patch?

-- 
Jens Axboe

Re: __io_file_supports_nowait for regular files

[Christoph Hellwig]

On Thu, Jul 21, 2022 at 09:59:40AM -0600, Jens Axboe wrote:
> On 7/21/22 9:37 AM, Christoph Hellwig wrote:
> > Hi Jens,
> > 
> > is there any good reason __io_file_supports_nowait checks the
> > blk_queue_nowait flag for regular files?  The FMODE_NOWAIT is set
> > for regular files that support nowait I/O, and should be all that
> > is needed.  Even with a block device that does not nonblocking
> > I/O some thing like reading from the page cache can be done
> > non-blocking.
> 
> Nope, we can probably kill that check then. Want to send a patch?

What would be a good way to verify it doesn't break any strange
assumptions?  I kinda don't fell too comfortable touching io_uring
guts.

> 
> -- 
> Jens Axboe
---end quoted text---

Re: __io_file_supports_nowait for regular files

[Jens Axboe]

On 7/21/22 10:23 AM, Christoph Hellwig wrote:
> On Thu, Jul 21, 2022 at 09:59:40AM -0600, Jens Axboe wrote:
>> On 7/21/22 9:37 AM, Christoph Hellwig wrote:
>>> Hi Jens,
>>>
>>> is there any good reason __io_file_supports_nowait checks the
>>> blk_queue_nowait flag for regular files?  The FMODE_NOWAIT is set
>>> for regular files that support nowait I/O, and should be all that
>>> is needed.  Even with a block device that does not nonblocking
>>> I/O some thing like reading from the page cache can be done
>>> non-blocking.
>>
>> Nope, we can probably kill that check then. Want to send a patch?
> 
> What would be a good way to verify it doesn't break any strange
> assumptions?  I kinda don't fell too comfortable touching io_uring
> guts.

Looking a bit deeper at this, FMODE_NOWAIT is about the file. The nowait
check for the bdev is about whether the driver honors NOWAIT
submissions. Any blk-mq driver will be fine, bio based ones probably
not. You could very well end up blocking off the submit path in that
case.

I'm not convinced we don't still need that check.

-- 
Jens Axboe

Re: __io_file_supports_nowait for regular files

[Christoph Hellwig]

On Thu, Jul 21, 2022 at 12:46:16PM -0600, Jens Axboe wrote:
> Looking a bit deeper at this, FMODE_NOWAIT is about the file.

Yes.

> The nowait
> check for the bdev is about whether the driver honors NOWAIT
> submissions. Any blk-mq driver will be fine, bio based ones probably
> not. You could very well end up blocking off the submit path in that
> case.

But do these submissions even matter for the high level interface?
We'd get -EAGAIN way more often without them (or all the time
for direct I/O), but does that strictly matter for the interface?

Note that for mny file systems (at least btrfs, f2fs and xfs) just
checking s_bdev is not enough any way as they can use multiple block
devices.

I'm also a little confused now that I'm looking more into this,
as iomap only uses REQ_NOWAIT for polled direct I/O to start with.
The legacy direct I/O code uses it for all writes as long as
IOCB_NOWAIT is set, so it seems like only the block device code
really makes extensive and most likely correct use of the
REQ_NOWAIT flag anyway.