On 6.8.2013 00:58, Yann E. MORIN wrote:
> On 2013-08-05 10:39 +0200, Michal Marek spake thusly:
>> Added Yann and the linux-kbuild list to CC. Reproducer:
>>
>> git checkout 1fe0135
>> make mrproper
>> make allmodconfig
>> make silentoldconfig
>> git checkout aa8032b
>> make allmodconfig
>> make silentoldconfig
>>
>> conf_write_autoconf() first calls conf_split_config() to generate the
>> include/config/**.h hierarchy, then generates include/config/auto.conf.
>> For some reason, conf_split_config() reads include/config/auto.conf,
>> which may not exist yet or may be out of date. Yann, can anything break
>> if we simply do not read that file from conf_split_config(), like this?
>>
>> diff --git a/scripts/kconfig/confdata.c b/scripts/kconfig/confdata.c
>> index c55c227..8c90835 100644
>> --- a/scripts/kconfig/confdata.c
>> +++ b/scripts/kconfig/confdata.c
>> @@ -829,16 +829,12 @@ next:
>>
>> static int conf_split_config(void)
>> {
>> - const char *name;
>> char path[PATH_MAX+1];
>> char *s, *d, c;
>> struct symbol *sym;
>> struct stat sb;
>> int res, i, fd;
>>
>> - name = conf_get_autoconfig_name();
>> - conf_read_simple(name, S_DEF_AUTO);
>> -
>> if (chdir("include/config"))
>> return 1;
[...]
>
> What I understand of the code is that, we're reading auto.conf with
> the S_DEF_AUTO flag, which is expected to set the SYMBOL_DEF_AUTO flags
> to symbols read from auto.conf.
>
> This is then used by conf_split_config to detect symbols that have
> changed.
Ah, _that_ is why it's there.
[...]
> The two following tests do not invlolve the HOTPLUG_PCI_.* symbols
> on purpose, and involve a pristine master c095ba7:
>
> $ make allmodconfig
> $ make silentoldconfig
> $ make silentoldconfig
> $ make silentoldconfig
>
> - Without Michal's proposed change, unaffected symbols are not
> touched. For example, mtime and ctime for zswap.h are not changed
> after the second and third silentoldconfig.
>
> - With Michal's proposed change, then mtime and ctime for zswap.h are
> changed after each silentoldconfig, even though CONFIG_ZSWAP was not
> changed.
Yes, I only did a simple diff -rq between two generated include/config
directories and did not think about mtimes.
> One solution would be to read auto.conf *before* we fuzz around symbols,
> not after. We should probably read auto.conf just after we read .config.
Maybe the safest solution would be to silence the warning if def ==
S_DEF_AUTO. Because auto.conf is only read by kconfig in
conf_split_config(), when it is expected to be out of date.
Michal
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
