Re: [PATCH RFC] xfs: add FALLOC_FL_WRITE_ZEROES to XFS code base

[lukas@herbolt.com]

On 2025-10-04 06:16, Christoph Hellwig wrote:
> On Fri, Oct 03, 2025 at 09:00:20PM -0700, Darrick J. Wong wrote:
>> > > -	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
>> > > +	if (mode & FALLOC_FL_WRITE_ZEROES) {
>> > > +		if (!bdev_write_zeroes_unmap_sectors(inode->i_sb->s_bdev))
>> 
>>      		                         not correct ^^^^^^^^^^^^^^^^^^^
>> 
>> You need to find the block device associated with the file because XFS
>> is a multi-device filesystem.
> 
> Indeed. xfs_inode_buftarg will do the work, but we'll need to ensure

Thanks for xfs_inode_buftarg pointer.

> the RT bit doesn't get flipped, i.e. it needs to hold a lock between
> that check and allocation the blocks if there were none yet.

I am having bit of trouble with that. If I get it right we should hold
the XFS_ILOCK_EXCL but this lock is then grabbed in the 
xfs_trans_alloc_inode.

So I would need to release before and there would be again a small 
window
where the RT flag can be flipped.

Looking at the xfs_alloc_file_space, there is also check for the RT bit 
without
lock, so this also need an attention.
     rt = XFS_IS_REALTIME_INODE(ip);
     extsz = xfs_get_extsz_hint(ip);

Or the xfs_trans_alloc_inode would need to check if we are a;ready 
holding the
lock Is there a way how to check the current thread is the owner of the
xfs_ilock rw_sem?

Something like?
---
static inline bool is_current_writer(struct rw_semaphore *sem)
{
     /* The rwsem_owner() helper can be used to get the owner */
     return sem->owner == current;
}
---