|
The core of the problem is an inconsistency in the MDT filesystem that caused the filesystem to be remounted read-only (-30 = -EROFS) to prevent cascading corruption. This will need a full e2fsck run to repair: e2fsck -fy /dev/<MDTDEV>. Please ensure you have the latest e2fsprogs-1.42.13.wc5 installed, which fixes a number of issues in older versions. You may want to consider doing a dry-run e2fsck -fn /dev/<MDTDEV> to log and review what problems are reported before starting the repair. If you want to be especially careful, you can create a snapshot of the MDT device, or do a full device-level backup (dd if=/dev/<MDTDEV> of=/dev/backup bs=1M) in case of problems.
The LBUG is the result of the read-only filesystem not being handled robustly, which is fixed by a patch http://review.whamcloud.com/19856 "LU-6696 llog: improve error handling", though it is not particularly beneficial to apply this patch to your system once the MDT problem is resolved, since by the time this assertion is hit the read-only MDT is not very useful anymore.
|
|
Looks like the e2fsck failed out with:
I've subsequently mounted it successfully albeit with warnings:
We're running some large rsyncs onto it now but I expect more issues of course.
|
|
Yang Sheng
Could you please assist with this issue?
Thanks
Peter
|
|
It looks like you will need to run "e2fsck -fy" again to repair the corrupted directory. I'm not sure why e2fsck aborted, I haven't seen a problem like that before.
|
|
The e2fsck seemed to stop at the same place again:
Any suggestions...?
|
|
We're quickly reaching the point where we're going to have to consider clearing things and starting from scratch which is something we'd really rather not do. We'd appreciate any other options you can present. If remote access would be useful, we can provide that.
If clearing things is the only real option here, can we provide any extra info to determine why this might have happened? As far as we can tell, the only thing that happened was the MDT filling up.
|
|
Stephen
How long would it take to get remote access in place? I know that you are based in the UK so what hours could a contact on site be available to work in realtime with one of our engineers?
Peter
|
|
We can get that in place. We're actually in Vancouver so Pacific hours. I'll set up some SSH and get back to you. Do you have an incoming IP I can restrict this to?
Please mail me on sfw@dneg.com - Thanks.
|
|
Stephen
I've sent you an email to get this going
Peter
|
|
Looking into this issue, it appears that the MDT filesystem is becoming full because of the use of hard-link trees for backup. This results in each file having a large link xattr that spills into an external block, as well as multiple directories referencing each file.
It would normally not be possible to have more than about 4500 bytes used per inode, even with the external xattr block, but the extra directory trees are consuming this space.
A file taken at random has a link count of 10 and an external xattr block (the File ACL block):
As for the e2fsck problem, I haven't been able to debug it yet because the problem takes several hours to hit. In the meantime, I've fixed the (first) problem that was causing the MDT to be remounted read-only:
I manually marked all of the blocks in group 0 used, and changed the group summary to match, as well as recomputed the block group checksum, which is the safest workaround given that I don't know which block(s) are actually in use, or which of those values is correct. It isn't clear if there are more errors like this, but I verified the next few groups had consistent block counts in the bitmap and group descriptors.
There is an e2fsck read-only check running under GDB to hopefully be able to debug the problem, but it will take about 8h to hit the point of the prior corruption.
|
|
Look into e2fsck code, It is failed at:
So looks like metadata is inconsistent. Maybe will passed after manual fixed.
|
|
The ext2fs_warn_bitmap() function is the source of the message Illegal inode number passed to ext2fs_test_inode_bitmap #0 for in-use inode map, but it turns out that this isn't the reason the e2fsck was aborted since in the first run this message appeared some time before e2fsck was aborted. That is caused by the later error Internal error: couldn't find dir_info for 3040644673, which was one of the inodes repaired in a previous run.
The Internal error message did appear for a different inode the first time, so it does seem possible that e2fsck is still repairing these inodes in each run, but not as well as they should be, causing an abort each run.
|
|
There were a couple of patches landed for 2.8.0 that make ldiskfs more robust in the face of similar corruption:
|
|
The problematic directory inode is 3040644673 as reported by e2fsck and hasn't been fixed after several runs:
Dumping the directory block shows nothing completely broken, except that the parent entry ".." is inode <2>, which is the ext4 root inode that should never be used by Lustre directories, but shouldn't cause any problems for e2fsck:
Even listing the directory shows nothing wrong with it except the ".." directory number:
The code that is triggering is in e2fsck/pass2.c::check_dir_block(), where e2fsck_dir_info_get_parent() fails:
if (e2fsck_dir_info_get_parent(ctx, dirent->inode,
&subdir_parent)) {
cd->pctx.ino = dirent->inode;
fix_problem(ctx, PR_2_NO_DIRINFO, &cd->pctx);
goto abort_free_dict;
}
after which e2fsck immediately aborts.
I'll need to dig into the population of the dir_info list to see why it is choking on "2" as the parent, since every directory under the root "/" directory also has <2> as the parent. Maybe that is special-cased in the code? Another possibility isn't that the parent directory "2" is the problem, but rather that this directory is not connected anywhere? I'd think that would be handled by adding it to lost+found or similar. It may be possible to debug this by creating a test filesystem that is similarly corrupted (non-root, maybe disconnected directory with <2> as the parent) to see what it is that triggers e2fsck to be unhappy.
|
|
I've tried to created a /mnt/lustre/dir1/dir2 which is disconnected, and e2fsck reports this
It does not complain anything in the pass2.
|
|
I tried to create a two-level subdirectory test/test2 and then zeroed out the leaf block of test1 so that there was no entry pointing to test2 and also set the ".." entry of test2 to <2>, but this didn't show anything different. Another test with just changing ".." of the test3 subdirectory to <2> didn't cause the problem.
I guess it is worthwhile to look into how directories are inserted into the parent directory list for e2fsck_dir_info_get_parent() to find. Something is preventing the 3040644673 inode from being found in the parent directory list.
|
|
So it seems like this was eventually resolved (we have not had further issues), but I do have a related question.
If we were to build this filesystem again, what would be the optimal inode size/ratio for performance considering we are likely to have 10+ hard links per file? Or does it not really matter and just ensuring we have more space available would be sufficient?
|
|
Daire, if you have a large number of links per file, then increasing the amount of space per inode would improve your space efficiency somewhat. Something like --mkfsoptions="-i 4608" would provide room for one external xattr block per file
|
|
Okay, thanks. We'll give that a try on our next filesystem. I wonder how that will interact with a RAID stripe of 4k too? But then we are using NVMe to back MDTs now so I doubt there is much RAID overhead any more.
|
|
The LBUG was fixed as part of LU-6696.
As for Lustre + 4KB RAID chunk size, I don't think the MDT will really affect performance much, since it is (presumably) configured as RAID-1+0 and not RAID-6.
|
|
Well, I was toying with the idea of using a 3 x NVMe (6.4TB) RAID5 for the MDT... We need lots of inodes and some redundancy. Maybe I should just use 4 cards and RAID10.
|
Generated at Sat Feb 10 02:15:56 UTC 2024 using Jira 9.4.14#940014-sha1:734e6822bbf0d45eff9af51f82432957f73aa32c.
dir.3040644673.bin
vmcore-dmesg.20160607.txt