On Fri, Aug 09, 2024 at 05:53:26PM +0200, Jürg Billeter wrote:
> Starting with 6.10, I'm seeing `copy_file_range()`, with source and
> destination being on the same FUSE filesystem[1], failing with EIO in
> some cases. The (low-level libfuse3) userspace filesystem does not
> implement `copy_file_range`, so the kernel falls back to the generic
> implementation. The userspace filesystem receives read requests and
> replies with the `FUSE_BUF_SPLICE_MOVE` flag.
>
> I'm not sure what exactly triggers the issue but it may depend on the
> file size, among other things. I can reproduce it fairly reliably
> attempting to copy files that are exactly 65536 bytes in size.
>
> 6.9 works fine but I see the issue in 6.10, 6.10.3 and also in current
> master ee9a43b7cfe2.
>
> 413e8f014c8b848e4ce939156f210df59fbd1c24 is the first bad commit
> commit 413e8f014c8b848e4ce939156f210df59fbd1c24 (HEAD)
> Author: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> Date: Sat Apr 20 03:50:06 2024 +0100
>
> fuse: Convert fuse_readpages_end() to use folio_end_read()
>
> Nobody checks the error flag on fuse folios, so stop setting it.
> Optimise the (optional) setting of the uptodate flag and clearing
> of the lock flag by using folio_end_read().
>
> Signed-off-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> Signed-off-by: Miklos Szeredi <mszeredi@xxxxxxxxxx>
>
> I've confirmed the bisection by reverting this commit on top of 6.10.3,
> which resolves the issue.
Umm. I don't see it. This is all that's in the commit:
for (i = 0; i < ap->num_pages; i++) {
- struct page *page = ap->pages[i];
+ struct folio *folio = page_folio(ap->pages[i]);
- if (!err)
- SetPageUptodate(page);
- else
- SetPageError(page);
- unlock_page(page);
- put_page(page);
+ folio_end_read(folio, !err);
+ folio_put(folio);
}
Do you have CONFIG_DEBUG_VM enabled? There are some debugging asserts
which that will enable that might indicate a problem.
Otherwise we can try splitting the commit into individual steps and
seeing which one shows the problem.
