lkml.org 
[lkml]   [2016]   [Jul]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [lkp] [ext4] 5405511e1a: ltp.acl_test01.fail]
Hi, Theodore

On Sun, Jul 10, 2016 at 11:15:53PM -0400, Theodore Ts'o wrote:
>On Mon, Jul 11, 2016 at 09:59:54AM +0800, kernel test robot wrote:
>>
>> FYI, we noticed the following commit:
>>
>> https://github.com/0day-ci/linux Vegard-Nossum/ext4-validate-number-of-clusters-in-group/20160708-041426
>> commit 5405511e1a984ab644fa9e29a0d3d958b835ab75 ("ext4: validate number of meta clusters in group")
>
>Hi, many thanks for the report. One question -- is there some way the
>reporting could be speeded up? Vegard's patch was sent late on the
>7th, and it looks like it hit the 0day-ci git repo a few hours later
>on the 8th, and was run early in the morning Chinese time. Was the
>delay until the 11th due to the weekend and the desire to have a human
>sanity check the result? Is that because there was a concern that
>there would be too many false positives?

Yes, you're right, although we have auto report feature (which has much
less delay) for high confidence errors/regressions detected, we still need
human sanity check for other cases to avoid too many false positives, it
will introduce some delays especially in weekend, sorry for that. We'll
improve our auto report part to cover more sensible cases.

>
>Also, if there was some way the 0day robot report could include a
>In-Reply-To: header to the patch being sent to the mailing list (in
>this case: <577EB740.10502@oracle.com>), it would make this even
>*more* useful.
>

Good point! We'll add this feature later.

Thanks,
Xiaolong

>
>Vegard, I'm guessing you didn't have a chance to test your patch
>before you sent it to the list?
>
> bit_max = ext4_num_clusters_in_group(sb, i);
> if ((bit_max >> 3) >= sb->s_blocksize) {
> ext4_msg(sb, KERN_WARNING, "clusters in "
> "group %u exceeds block size", i);
> goto failed_mount;
> }
>
>
>This is the test which is failing, but it will fail by default on
>pretty much all ext4 file systems, since by default there will be
>32768 blocks (clusters) per group, with a 4k block size (and 32768 >>
>3 == 4096). And in the test that failed, this was a 1k block size
>with 8192 blocks per blocks (and 8192 >> 3 == 1024).
>
>Anyway, as I mentioned before, I'd much rather do very specific sanity
>checking on superblock fields, instead of sanity checking calculated
>values such as ext4_num_clusters_in_group().
>
>Perhaps the easist thing to do is to run e2fsck -n on those file
>systems that are causing problems?
>
>Cheers,
>
> - Ted

\
 
 \ /
  Last update: 2016-07-11 07:41    [W:0.057 / U:8.768 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site