On Tue, Oct 18, 2022 at 04:24:46PM +0800, Hsin-Yi Wang wrote:
> Hi Jintao,
> 
> Thanks for pointing out and sorry for missing this. Does the following
> diff improve the issue?
> 
> diff --git a/fs/squashfs/file.c b/fs/squashfs/file.c
> index e56510964b229..7759bd70dfbf2 100644
> --- a/fs/squashfs/file.c
> +++ b/fs/squashfs/file.c
> @@ -600,7 +600,7 @@ static void squashfs_readahead(struct
> readahead_control *ractl)
> 
>                         /* Last page (if present) may have trailing
> bytes not filled */
>                         bytes = res % PAGE_SIZE;
> -                       if (pages[nr_pages - 1]->index == file_end && bytes)
> +                       if ((pages[nr_pages - 1]->index >> shift) ==
> file_end && bytes)
>                                 memzero_page(pages[nr_pages - 1], bytes,
>                                              PAGE_SIZE - bytes);
> 
> 
> readahead only handles the case that the first page and the last page
> have the same block index:
>     index = pages[0]->index >> shift;
>     if ((pages[nr_pages - 1]->index >> shift) != index)
>         goto skip_pages;
> 
> The diff above makes a difference to SQUASHFS_INVALID_BLK case, which
> will not be handled by squashfs_readahead_fragment() if
> index==file_end.
> With the above diff, it will now be memzero_page().

Hi Hsin-Yi Wang, thanks for the proposed diff. However, it was wrapped,
so I have to manually type the changes.

I compiled 6.1-rc1 with your diff applied. For testing, I have both
hello-world and lxd snaps installed. No problems on running the former.
On the latter, I got coredump when trying to start lxd services with
`snap start lxd`. The coredump for lxd processes are attached.

From above, I think b09a7a036d2035 ("squashfs: support reading fragments in readahead call") should be reverted until we come up with proper solution.

Thanks.

-- 
An old man doll... just what I always wanted! - Clara