[PATCH] xfs: fix the judgment of whether the file already has extents

[PATCH] xfs: fix the judgment of whether the file already has extents

[alexjlzheng]

From: Jinliang Zheng <alexjlzheng@tencent.com>

When we call create(), lseek() and write() sequentially, offset != 0
cannot be used as a judgment condition for whether the file already
has extents.

This patch uses prev.br_startoff instead of offset != 0.

Signed-off-by: Jinliang Zheng <alexjlzheng@tencent.com>
---
 fs/xfs/libxfs/xfs_bmap.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/xfs/libxfs/xfs_bmap.c b/fs/xfs/libxfs/xfs_bmap.c
index 36dd08d13293..94e7aeed9e95 100644
--- a/fs/xfs/libxfs/xfs_bmap.c
+++ b/fs/xfs/libxfs/xfs_bmap.c
@@ -3536,7 +3536,7 @@ xfs_bmap_btalloc_at_eof(
 	 * or it's the first allocation in a file, just try for a stripe aligned
 	 * allocation.
 	 */
-	if (ap->offset) {
+	if (ap->prev.br_startoff != NULLFILEOFF) {
 		xfs_extlen_t	nextminlen = 0;
 
 		/*
-- 
2.41.1

Re: [PATCH] xfs: fix the judgment of whether the file already has extents

[Christoph Hellwig]

On Sun, Oct 27, 2024 at 02:01:16AM +0800, alexjlzheng@gmail.com wrote:
> From: Jinliang Zheng <alexjlzheng@tencent.com>
> 
> When we call create(), lseek() and write() sequentially, offset != 0
> cannot be used as a judgment condition for whether the file already
> has extents.
> 
> This patch uses prev.br_startoff instead of offset != 0.

This changed the predicate from "are we at offset 0" to "are there
any allocations before that".  That's a pretty big semantic change.
Maybe a good one, maybe not.  Can you explain what workload it helps
you with?

Re: [PATCH] xfs: fix the judgment of whether the file already has extents

[Jinliang Zheng]

On Mon, 28 Oct 2024 02:41:01 -0700, hch@infradead.org wrote:
> On Sun, Oct 27, 2024 at 02:01:16AM +0800, alexjlzheng@gmail.com wrote:
> > From: Jinliang Zheng <alexjlzheng@tencent.com>
> > 
> > When we call create(), lseek() and write() sequentially, offset != 0
> > cannot be used as a judgment condition for whether the file already
> > has extents.
> > 
> > This patch uses prev.br_startoff instead of offset != 0.
>
> This changed the predicate from "are we at offset 0" to "are there
> any allocations before that".  That's a pretty big semantic change.
> Maybe a good one, maybe not.  Can you explain what workload it helps
> you with?


Thanks for your reply.

I noticed this because I was confused when reading the code here. The code
comment here says:

/*
 * If there are already extents in the file, try an exact EOF block
 * allocation to extend the file as a contiguous extent. If that fails,
 * or it's the first allocation in a file, just try for a stripe aligned
 * allocation.
 */

But as you said, the semantics of the current code is "are we at offset 0",
not "are there any allocations before that".

Therefore, I think it is better to use "prev.br_startoff != NULLFILEOFF"
instead of the current "offset != 0", at least its semantics are more
consistent with the intention in the code comment and reduce confusion.

But if the semantics here have indeed changed to the point where it is
inconsistent with the code comment, my suggestion is to update the code
comment here.

Thank you. :)
Jinliang Zheng

Re: [PATCH] xfs: fix the judgment of whether the file already has extents

[Jinliang Zheng]

On Mon, 28 Oct 2024 18:33:32 +0800, alexjlzheng@tencent.com wrote:
> On Mon, 28 Oct 2024 02:41:01 -0700, hch@infradead.org wrote:
> > On Sun, Oct 27, 2024 at 02:01:16AM +0800, alexjlzheng@gmail.com wrote:
> > > From: Jinliang Zheng <alexjlzheng@tencent.com>
> > > 
> > > When we call create(), lseek() and write() sequentially, offset != 0
> > > cannot be used as a judgment condition for whether the file already
> > > has extents.
> > > 
> > > This patch uses prev.br_startoff instead of offset != 0.
> >
> > This changed the predicate from "are we at offset 0" to "are there
> > any allocations before that".  That's a pretty big semantic change.
> > Maybe a good one, maybe not.  Can you explain what workload it helps
> > you with?
> 
> 
> Thanks for your reply.
> 
> I noticed this because I was confused when reading the code here. The code
> comment here says:
> 
> /*
>  * If there are already extents in the file, try an exact EOF block
>  * allocation to extend the file as a contiguous extent. If that fails,
>  * or it's the first allocation in a file, just try for a stripe aligned
>  * allocation.
>  */
> 
> But as you said, the semantics of the current code is "are we at offset 0",
> not "are there any allocations before that".

By the way, we only get here if got is or after EOF, so "are there any allocations
before that" means "are there already extents in the file".

Thank you, again. :)
Jinliang Zheng

> 
> Therefore, I think it is better to use "prev.br_startoff != NULLFILEOFF"
> instead of the current "offset != 0", at least its semantics are more
> consistent with the intention in the code comment and reduce confusion.
> 
> But if the semantics here have indeed changed to the point where it is
> inconsistent with the code comment, my suggestion is to update the code
> comment here.
> 
> Thank you. :)
> Jinliang Zheng

Re: [PATCH] xfs: fix the judgment of whether the file already has extents

[Dave Chinner]

On Sun, Oct 27, 2024 at 02:01:16AM +0800, alexjlzheng@gmail.com wrote:
> From: Jinliang Zheng <alexjlzheng@tencent.com>
> 
> When we call create(), lseek() and write() sequentially, offset != 0
> cannot be used as a judgment condition for whether the file already
> has extents.
> 
> This patch uses prev.br_startoff instead of offset != 0.
> 
> Signed-off-by: Jinliang Zheng <alexjlzheng@tencent.com>
> ---
>  fs/xfs/libxfs/xfs_bmap.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/fs/xfs/libxfs/xfs_bmap.c b/fs/xfs/libxfs/xfs_bmap.c
> index 36dd08d13293..94e7aeed9e95 100644
> --- a/fs/xfs/libxfs/xfs_bmap.c
> +++ b/fs/xfs/libxfs/xfs_bmap.c
> @@ -3536,7 +3536,7 @@ xfs_bmap_btalloc_at_eof(
>  	 * or it's the first allocation in a file, just try for a stripe aligned
>  	 * allocation.
>  	 */
> -	if (ap->offset) {
> +	if (ap->prev.br_startoff != NULLFILEOFF) {
>  		xfs_extlen_t	nextminlen = 0;

Makes sense, but the logic is not correct. See xfs_bmap_adjacent()
on how it sets up the ap->blkno target for exact eof bno allocation.

-Dave.
-- 
Dave Chinner
david@fromorbit.com