Re: [PATCH v2 2/10] xfs: Add support FALLOC_FL_INSERT_RANGE for fallocate

[Brian Foster]

On Mon, May 12, 2014 at 06:42:37PM +0900, Namjae Jeon wrote:
> 
> > > +xfs_bmap_split_extent(
> > > +	struct xfs_inode	*ip,
> > > +	xfs_fileoff_t		split_fsb,
> > > +	xfs_extnum_t		*split_ext)
> > > +{
> > > +	struct xfs_mount        *mp = ip->i_mount;
> > > +	struct xfs_trans	*tp;
> > > +	struct xfs_bmap_free	free_list;
> > > +	xfs_fsblock_t		firstfsb;
> > > +	int			committed;
> > > +	int			error;
> > > +
> > > +	tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
> > > +	error = xfs_trans_reserve(tp, &M_RES(mp)->tr_write,
> > > +			XFS_DIOSTRAT_SPACE_RES(mp, 0), 0);
> > > +
> > > +	if (error) {
> > > +		/*
> > > +		 * Free the transaction structure.
> > > +		 */
> > > +		ASSERT(XFS_FORCED_SHUTDOWN(mp));
> > 
> Hi, Brian.
> > As in the other patch, we're attempting to reserve fs blocks for the
> > transaction, so ENOSPC is a possibility that I think the assert should
> > accommodate.
> How about removing the ASSERT completely as suggessted by Dave
> in other thread?
> 

Yeah, that works too. If Dave prefers to just remove these asserts
that's fine with me. I just wanted to make sure we aren't adding
spurious asserts for valid failures.

> > 
...
> > > diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> > > index 97855c5..392b029 100644
> > > --- a/fs/xfs/xfs_file.c
> > > +++ b/fs/xfs/xfs_file.c
> > > @@ -760,7 +760,8 @@ xfs_file_fallocate(
> > >  	if (!S_ISREG(inode->i_mode))
> > >  		return -EINVAL;
> > >  	if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
> > > -		     FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE))
> > > +		     FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |
> > > +		     FALLOC_FL_INSERT_RANGE))
> > >  		return -EOPNOTSUPP;
> > >
> > >  	xfs_ilock(ip, XFS_IOLOCK_EXCL);
> > > @@ -790,6 +791,40 @@ xfs_file_fallocate(
> > >  		error = xfs_collapse_file_space(ip, offset, len);
> > >  		if (error)
> > >  			goto out_unlock;
> > > +	} else if (mode & FALLOC_FL_INSERT_RANGE) {
> > > +		unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
> > > +		struct iattr iattr;
> > > +
> > > +		if (offset & blksize_mask || len & blksize_mask) {
> > > +			error = -EINVAL;
> > > +			goto out_unlock;
> > > +		}
> > > +
> > > +		/* Check for wrap through zero */
> > > +		if (inode->i_size + len > inode->i_sb->s_maxbytes) {
> > > +			error = -EFBIG;
> > > +			goto out_unlock;
> > > +		}
> > > +
> > > +		/* Offset should be less than i_size */
> > > +		if (offset >= i_size_read(inode)) {
> > > +			error = -EINVAL;
> > > +			goto out_unlock;
> > > +		}
> > > +
> > > +		/*
> > > +		 * The first thing we do is to expand file to
> > > +		 * avoid data loss if there is error while shifting
> > > +		 */
> > > +		iattr.ia_valid = ATTR_SIZE;
> > > +		iattr.ia_size = i_size_read(inode) + len;
> > > +		error = xfs_setattr_size(ip, &iattr);
> > > +		if (error)
> > > +			goto out_unlock;
> > 
> > I don't necessarily know that it's problematic to do the setattr before
> > the bmap fixup. We'll have a chance for partial completion of this
> > operation either way. But I'm not a fan of the code duplication here.
> > This also still skips the time update in the event of insert space
> > failure, though perhaps that's not such a big deal if we're returning an
> > error.
> > 
> > I think it would be better to leave things organized as before and
> > introduce an error2 variable and a &nrshifts or some such parameter to
> > xfs_insert_file_space() that initializes to 0 and returns the number of
> > record shifts. The caller can then decide whether it's appropriate to
> > break out immediately or do the inode size update and return the error.
> > 
> > Perhaps not the cleanest thing in the world, but also not the first
> > place we would use 'error2' to manage error priorities (grep around for
> > it)...
> Yes, Right. I also thought such sequence at first. But we should consider
> sudden power off and unplug device case during shifting extent.
> While we are in the middle of shifitng extents and if there is sudden
> power failure user can still think that data is lost as we won't get any
> chance to update the file size in these cases.
> Updating file size before the shifitng operation can start will prevent this.
> 
> Thanks.

Hmm, fair point. That seems less critical to me than the general error
sequence, but if we want to handle that case I think we could still fix
the duplication in xfs_file_fallocate(). We could possibly factor out
the common bits (update time and set size) into a helper, or what seems
a bit cleaner on first thought, move the bulk of the (mode &
FALLOC_FL_INSERT_RANGE) block to after the common part. Then the
function looks something like this:

	...
	xfs_ilock();

	/* pre-inode fixup ops */
	if (mode & ...) {
		...
	} else if (mode & FALLOC_FL_INSERT_RANGE) {
		/* comment as to what's going on here :) */

		/* error checks */

		new_size = ...;
		do_file_insert = 1;
	}
	...
	xfs_trans_ichgtime();
	xfs_setattr_size();
	...

	/*
	 * Some operations are performed after the inode size is updated. For
	 * example, insert range expands the address space of the file, shifts
	 * all subsequent extents over and allocates space into the hole.
	 * Updating the size first ensures that shifted extents aren't left
	 * hanging past EOF in the event of a crash or failure.
	 */
	if (do_file_insert) {
		/* alloc space */
		...
	}
	...

That seems a bit cleaner to me, but I'm not wedded to it. Thoughts? It
might be worth soliciting some other thoughts/ideas before reworking it.
Thanks.

Brian

> > 
> > Brian
> > 
> > > +
> > > +		error = xfs_insert_file_space(ip, offset, len);
> > > +		if (error)
> > > +			goto out_unlock;
> > >  	} else {
> > >  		if (!(mode & FALLOC_FL_KEEP_SIZE) &&
> > >  		    offset + len > i_size_read(inode)) {
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/