Re: [PATCH] loop: stop using vfs_iter_{read,write} for buffered I/O

[Christoph Hellwig <hch@lst.de>]

On Wed, Apr 09, 2025 at 09:44:41PM +0800, Ming Lei wrote:
> On Wed, Apr 09, 2025 at 03:09:40PM +0200, Christoph Hellwig wrote:
> > vfs_iter_{read,write} always perform direct I/O when the file has the
> > O_DIRECT flag set, which breaks disabling direct I/O using the
> > LOOP_SET_STATUS / LOOP_SET_STATUS64 ioctls.
> 
> So dio is disabled automatically because lo_offset is changed in
> LOOP_SET_STATUS, but backing file is still opened with O_DIRECT,
> then dio fails?
> 
> But Darrick reports it is caused by changing sector size, instead of
> LOOP_SET_STATUS.

LOOP_SET_STATUS changes the direct I/O flag.

This is the minimal reproducer, dev needs to be a 4k lba size device:

dev=/dev/nvme0n1

mkfs.xfs -f $dev
mount $dev /mnt

truncate -s 30g /mnt/a
losetup --direct-io=on -f --show /mnt/a
losetup --direct-io=off /dev/loop0
losetup --sector-size 2048 /dev/loop0
mkfs.xfs /dev/loop0

mkfs then fails with an I/O error.

(I plan to wire up something like this for blktests)

> > This was recenly reported as a regression, but as far as I can tell
> > was only uncovered by better checking for block sizes and has been
> > around since the direct I/O support was added.
> 
> What is the 1st real bad commit for this regression? I think it is useful
> for backporting. Or it is new test case?

Not entirely sure, maybe Darrick can fill in.

> 
> > 
> > Fix this by using the existing aio code that calls the raw read/write
> > iter methods instead.  Note that despite the comments there is no need
> > for block drivers to ever call flush_dcache_page themselves, and the
> > call is a left-over from prehistoric times.
> > 
> > Fixes: ab1cb278bc70 ("block: loop: introduce ioctl command of LOOP_SET_DIRECT_IO")
> 
> Why is the issue related with ioctl(LOOP_SET_DIRECT_IO)?
> 
> 
> Thanks, 
> Ming
---end quoted text---