[PATCH] makedumpfile: make reserve_diskspace do nothing for flattened format

[HAGIO KAZUHITO(萩尾 一仁)]

Hi Jiri,

sorry for the long delay.

On 2024/06/19 21:04, Jiri Bohac wrote:
> makedumpfile: make reserve_diskspace do nothing for flattened format
> 
> reserve_diskspace() is called by write_elf_header() to make sure there is
> always space to write the program header, even if writing other data fails
> because of ENOSPC.
> 
> This is harmful when writing the flattened format to STDOUT for two reasons:
> 
> First, it actually wastes disk space, because first the block of zeroes is sent
> to STDOUT by reserve_diskspace() and then the actual program header is sent,
> meant to overwrite the zeroes when the flattened format is rearranged.
> 
> Second, the algorithm used to read flattened format directly by the crash
> program does not cope with the flattened file containing two chunks meant for
> the same offset. It uses a binary search on a sorted array of flat_data headers
> to find the data in the flat file. It may return the zeroed chunk written by
> reserve_diskspace() near the beginning of the file instead of the actual ELF
> header located near the end of the flattened file.

Thank you for the patch, I found a vmcore that reproduced the issue:

$ makedumpfile -FEd 31 vmcore > dump.FEd31
$ crash vmlinux dump.FEd31
...
realloc: No such file or directory
cannot realloc resized ELF header buffer
$

and the patch fixed this, so

Acked-by: Kazuhito Hagio <k-hagio-ab at nec.com>

(Masa will apply the patch, please wait for a while.)

Thanks,
Kazu


> 
> Fixes: e39216fce9f73759509ec158e39c289e6c211125 ("Make the incomplete dumpfile generated by ENOSPC error analyzable.")
> Signed-off-by: Jiri Bohac <jbohac at suse.cz>
> 
> ---
>   makedumpfile.c | 3 +++
>   1 file changed, 3 insertions(+)
> 
> diff --git a/makedumpfile.c b/makedumpfile.c
> index cadc596..9624c3f 100644
> --- a/makedumpfile.c
> +++ b/makedumpfile.c
> @@ -5206,6 +5206,9 @@ reserve_diskspace(int fd, off_t start_offset, off_t end_offset, char *file_name)
>   
>   	int ret = FALSE;
>   
> +	if (info->flag_flatten)
> +		return TRUE;
> +
>   	assert(start_offset < end_offset);
>   	buf_size = end_offset - start_offset;
>