Thread (1 message) 1 message, 1 author, 2021-07-18

Re: Read time tree block corruption detected

From: Qu Wenruo <hidden>
Date: 2021-07-18 09:32:30


On 2021/7/18 下午4:46, pepperpoint@mb.ardentcoding.com wrote:
Hi Qu,

When I find which directory some of the filenames are located, they are under /var/lib.
I had subvolume for /var which was created probably around 2018.
Then it's possible by somehow we allowed that hardlink to directory.

Not sure if it's a bug in VFS layer or in btrfs itself.

But around 2019 (aka, v5.2 kernel), that check for refs of directory is
introduced and at the same time, write-time tree checker is introduced.

This means if the bug happens after v5.2 kernel, it will be rejected
before submitting to disk.

So the problem definitely happens before the install of v5.2 kernel.
I don't remember how I created this but I probably use rsync to copy the files from existing /var
or created a snapshot of root and delete other files that is not under /var.
But that's still pretty weird.
Around June, I tried to move the filesystem to another partition through btrfs device add and btrfs device remove
but failed due to that error and was advised to use btrfs replace instead.

Then at the beginning of this month, I reorganized it merging most of the /var content back to root
and created subvolume for /var/lib/mysql and /var/lib/mongodb. I encountered an error when
I copy some of the files through cp --reflink but I failed for /var/lib/mysql so I created a snapshot
from /var and remove the extra files.

This is also the time I saw the errors in the log. Before that, the errors was not in the log.
At least, we should prevent such problem from reaching disk.

If you reverted to older LTS kernel, using Arch Linux Archive, it would
be possible to continue deleting the subvolume and solve the problem.

After the root 363 get fully deleted, you can verify that tree block get
deleted by the following command:

# btrfs ins dump-tree -t extent <device> | grep 174113599488 -A 3

Which should show no output.

Thanks,
Qu
Regards,
Lester


‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐

On Sunday, July 18th, 2021 at 3:27 PM, Qu Wenruo [off-list ref] wrote:
quoted
Hi,

BTW, it's really important for us to know how the directory is hardlinked.

Thus I salvaged the filenames found in the half-dropped root 363.

Since it may contain confidential info, I send the filename list to you

off-list.

If you can remind what the root 363 is used for, and any possible

operations which may be involved in that subvolume, it's better to reply

it to the mail list so that we can get some clue on the root cause.

Thanks,

Qu

On 2021/7/18 下午3:15, Qu Wenruo wrote:
quoted
On 2021/7/18 下午1:26, pepperpoint@mb.ardentcoding.com wrote:
quoted
Hi Qu,

May I know if there are any leads on this? What should I do for now?
Sorry for the late reply.

With the image dump, it's much easier to find what's going wrong.

-   About root 363

       It's an orphan root, thus user can't access it directly.

Furthermore, it's being dropped, thus "btrfs ins dump-tree -t 363"

   reports transid error, as part of the tree has already been dropped,

   and this is expected.

So far your fs is still OK, except that reported error.

-   About the offending tree block

       The offending tree block only belongs to the delete subvolume 363,

       thus it should be delete soon.

But unfortunately due to the corrupted content, it's unable to be

   deleted.

For now, if you can re-compile btrfs module, we can workaround the

   problem by temporarily disable read-time tree-checker so that the

   deletion can continue, and after the root 363 get fully deleted, the

   problem should be gone.

Or you can use older kernel, any kernel <= v5.1 should not have the

   enhanced check, thus can continue with the subvolume deletion.

If you want to go through the re-compile path, the needed diff is

   attached

Thanks,

Qu
quoted
Regards,

Lester

‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐

On Saturday, July 17th, 2021 at 8:51 PM,

pepperpoint@mb.ardentcoding.com wrote:
quoted
Hi Qu,

I run btrfs ins dump-tree -t 363 unmounted but the same error

appears. Rerunning btrfs check does not show any error.

Regards,

Lester

‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐

On Saturday, July 17th, 2021 at 6:48 PM, Qu Wenruo -

quwenruo.btrfs@gmx.com wrote:
quoted
On 2021/7/17 下午6:34, pepperpoint@mb.ardentcoding.com wrote:
quoted
Hi Qu,

Unfortunately I cannot find subvolume 363

btrfs subvolume list /run/media/root
====================================

ID 361 gen 1814826 top level 584 path @/live/snapshot

ID 364 gen 1814414 top level 5 path @vtmp/live/snapshot

ID 369 gen 1814414 top level 5 path @vlmachines/live/snapshot

ID 493 gen 1814414 top level 5 path @vlportables/live/snapshot

ID 579 gen 1814828 top level 5 path @vlog/live/snapshot

ID 580 gen 1814414 top level 5 path @vcache/live/snapshot

ID 581 gen 1814414 top level 5 path @vlmongodb/live/snapshot

ID 582 gen 1814414 top level 5 path @vlmysql/live/snapshot

ID 583 gen 1814414 top level 5 path @vspool/live/snapshot

ID 584 gen 1814414 top level 5 path @

ID 598 gen 1813420 top level 584 path @/4/snapshot
Maybe 363 is some subvolume get deleted and later snapshot of it still

exists.

This will be harder to debug.

Can you take a btrfs-image dump of your filesystem? (needs to be taken

with the fs unmounted).

The dumped image will contain your metadata, including file names and

directory structures, but no data inside those files.

Although btrfs-image has "-s" option to mask the filenames, but

considering the filename in this case is useful to locate the inode, I

guess it's better to take the image without any "-s" option.
quoted
btrfs ins dump-tree -t 363 /dev/dm-0 | grep -A 5 "(286 "
========================================================

parent transid verify failed on 174170742784 wanted 1789655 found

1812621

parent transid verify failed on 174170742784 wanted 1789655 found

1812621

parent transid verify failed on 174170742784 wanted 1789655 found

1812621

Ignoring transid failure

ERROR: child eb corrupted: parent bytenr=174170738688 item=0 parent

level=2 child bytenr=174170742784 child level=0
This transid mismatch may be a problem when running dump-tree on

mounted

fs, since you mentioned btrfs check reported no error, there shouldn't

be a transid mismatch error.

Anyway, if you can upload the btrfs-image dump, it would be much easier

for us to debug and find out what's really going.

Thanks,

Qu
quoted
‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐

On Saturday, July 17th, 2021 at 6:12 PM, Qu Wenruo wqu@suse.com wrote:
quoted
On 2021/7/17 下午4:57, pepperpoint@mb.ardentcoding.com wrote:
quoted
Hi Qu,

I don't know how the directory was created but last month, I used

btrfs device add and btrfs device remove to move the filesystem

from one partition to another. It failed because of the same

error and was advised to use btrfs replace instead. I don't know

if the error also happened before I move the file system as I

don't have any previous logs.
It definitely happens before you moving the fs.

As regular dev replacing/add/move only relocates the metadata, but

not

touching the fs trees.
quoted
Here is the result when I search for the inodes you mentioned if

it helps:

find /run/media/root -inum 260 -exec ls -ldi {} \;
==================================================

260 -rw-r--r-- 1 root root 36864 Jun 25 06:22

/run/media/root/@vcache/live/snapshot/app-info/cache/en_US.cache

260 drwx------ 1 mongodb mongodb 136 Sep 12 2020

/run/media/root/@vlmongodb/live/snapshot/diagnostic.data

260 -rw-rw---- 1 mysql mysql 50331648 Sep 13 2015

/run/media/root/@vlmysql/live/snapshot/ib_logfile0

260 -rw-r----- 1 root lp 8641 Mar 5 2014

/run/media/root/@vspool/live/snapshot/cups/d00001-001

260 dr-xr-xr-x 1 root root 0 Sep 13 2013

/run/media/root/@/live/snapshot/sys

260 dr-xr-xr-x 1 root root 0 Sep 13 2013

/run/media/root/@/4/snapshot/sys
Since btrfs can have the same inode number inside different

subvolumes,

you may want to limit the search inside subvolume 363.

"-mount" option of find can do that, you only need to locate

subvolume

363 by "btrfs subv list".

But from these output I guess above two "sys" directory are more

possible.

Is there any directory named "blaklight" inside those directory?
quoted
find /run/media/root -inum 286 -exec ls -ldi {} \;
==================================================

286 -rw-r--r-- 1 root root 96 Aug 16 2015

/run/media/root/@vcache/live/snapshot/fontconfig/4b172ca7f111e3cffadc3636415fead9-le64.cache-4

286 -rw-rw---- 1 mysql mysql 4096 Sep 15 2013

/run/media/root/@vlmysql/live/snapshot/mysql/columns_priv.MYI

286 -rw-r-----+ 1 root systemd-journal 16777216 Jul 4 01:14

/run/media/root/@vlog/live/snapshot/journal/5098dd7845ae46d3ba1826c68a809a7c/user-1000@fbd9f65d0ea349f6b996716280e6c4dd-00000000002314c5-0005c5cb84a3a438.journal
This is interesting, it means the inode 286 is not accessible.

It can be some orphan inode, but would you dump subvolume 363 then

try

to locate the inode 286?

One example command would be:

btrfs ins dump-tree -t 363 <dev> | grep -A 5 "(286 "
====================================================

Thanks,

Qu
quoted
Directories with pattern /root/@<dir>/live/snapshot/ are

subvolumes and directories with pattern

/root/@<dir>/<num>/snapshot/ are snapshots of live.

‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐

On Saturday, July 17th, 2021 at 4:14 PM, Qu Wenruo

quwenruo.btrfs@gmx.com wrote:
quoted
On 2021/7/17 下午3:51, pepperpoint@mb.ardentcoding.com wrote:
quoted
Hi Qu,

Please see below for the dump.

btrfs-progs v5.12.1

leaf 174113599488 items 18 free space 2008 generation 1330906

owner 363

leaf 174113599488 flags 0x1(WRITTEN) backref revision 1

fs uuid a7d327c4-8594-4116-a6f8-8aa2a4162063

chunk uuid f885f49e-14a0-4c80-9c12-c2302b9a0229

item 0 key (5471 INODE_ITEM 0) itemoff 3835 itemsize 160

generation 2063 transid 27726 size 40 nbytes 40

block group 0 mode 100600 links 1 uid 0 gid 100 rdev 0

sequence 1501 flags 0x0(none)

atime 1386484844.468769570 (2013-12-08 14:40:44)

ctime 1386484844.468769570 (2013-12-08 14:40:44)

mtime 1386484844.468769570 (2013-12-08 14:40:44)

otime 0.0 (1970-01-01 08:00:00)

item 1 key (5471 INODE_REF 4399) itemoff 3824 itemsize 11

index 12 namelen 1 name: 8

item 2 key (5471 EXTENT_DATA 0) itemoff 3763 itemsize 61

generation 27726 type 0 (inline)

inline extent data size 40 ram_bytes 40 compression 0 (none)

item 3 key (5645 INODE_ITEM 0) itemoff 3603 itemsize 160

generation 2542 transid 61261 size 40 nbytes 40

block group 0 mode 100600 links 1 uid 0 gid 100 rdev 0

sequence 24769 flags 0x0(none)

atime 1394335806.351857522 (2014-03-09 11:30:06)

ctime 1394335827.344389955 (2014-03-09 11:30:27)

mtime 1394335827.344389955 (2014-03-09 11:30:27)

otime 0.0 (1970-01-01 08:00:00)

item 4 key (5645 INODE_REF 4399) itemoff 3592 itemsize 11

index 13 namelen 1 name: 7

item 5 key (5645 EXTENT_DATA 0) itemoff 3531 itemsize 61

generation 61261 type 0 (inline)

inline extent data size 40 ram_bytes 40 compression 0 (none)

item 6 key (7222 INODE_ITEM 0) itemoff 3371 itemsize 160

generation 5754 transid 5767 size 307 nbytes 307

block group 0 mode 100644 links 1 uid 0 gid 0 rdev 0

sequence 7 flags 0x0(none)

atime 1379834835.428558020 (2013-09-22 15:27:15)

ctime 1379834835.428558020 (2013-09-22 15:27:15)

mtime 1379834835.428558020 (2013-09-22 15:27:15)

otime 0.0 (1970-01-01 08:00:00)

item 7 key (7222 INODE_REF 287) itemoff 3344 itemsize 27

index 6 namelen 17 name: dhcpcd-eth0.lease

item 8 key (7222 EXTENT_DATA 0) itemoff 3016 itemsize 328

generation 5767 type 0 (inline)

inline extent data size 307 ram_bytes 307 compression 0 (none)

item 9 key (7415 INODE_ITEM 0) itemoff 2856 itemsize 160

generation 5904 transid 1330906 size 180 nbytes 0

block group 0 mode 40755 links 2 uid 0 gid 0 rdev 0

sequence 177 flags 0x0(none)

atime 1483277713.141980592 (2017-01-01 21:35:13)

ctime 1563162901.234656246 (2019-07-15 11:55:01)

mtime 1406534032.158605559 (2014-07-28 15:53:52)

otime 0.0 (1970-01-01 08:00:00)
This inode is indeed a directory.

But it has two hard links, which is definitely something

unexpected.

Under Linux we shouldn't have any hardlink for directory, as it

would

easily lead to loops.
quoted
item 10 key (7415 INODE_REF 260) itemoff 2837 itemsize 19

index 28 namelen 9 name: backlight
Its parent inode is 260 in the same root, with the name backlight.
quoted
item 11 key (7415 INODE_REF 286) itemoff 2818 itemsize 19

index 3 namelen 9 name: backlight
Another hardlink in inode 286, which is definitely a regular thing.

Btrfs-progs lacks the ability to detect such problem, we need to

enhance

it first.

But do you have any idea how this directory get created?

It looks like the content of sysfs.

Thanks,

Qu
quoted
item 12 key (7415 DIR_ITEM 3128336373) itemoff 2746 itemsize 72

location key (120417 INODE_ITEM 0) type FILE

transid 117279 data_len 0 name_len 42

name: pci-0000:00:02.0:backlight:intel_backlight

item 13 key (7415 DIR_ITEM 3218198317) itemoff 2705 itemsize 41

location key (7487 INODE_ITEM 0) type FILE

transid 5992 data_len 0 name_len 11

name: acpi_video0

item 14 key (7415 DIR_ITEM 3582254411) itemoff 2638 itemsize 67

location key (55325 INODE_ITEM 0) type FILE

transid 63351 data_len 0 name_len 37

name: platform-VPC2004:00:backlight:ideapad

item 15 key (7415 DIR_INDEX 2) itemoff 2597 itemsize 41

location key (7487 INODE_ITEM 0) type FILE

transid 5992 data_len 0 name_len 11

name: acpi_video0

item 16 key (7415 DIR_INDEX 4) itemoff 2530 itemsize 67

location key (55325 INODE_ITEM 0) type FILE

transid 63351 data_len 0 name_len 37

name: platform-VPC2004:00:backlight:ideapad

item 17 key (7415 DIR_INDEX 5) itemoff 2458 itemsize 72

location key (120417 INODE_ITEM 0) type FILE

transid 117279 data_len 0 name_len 42

name: pci-0000:00:02.0:backlight:intel_backlight

‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐

On Saturday, July 17th, 2021 at 3:05 PM, Qu Wenruo

quwenruo.btrfs@gmx.com wrote:
quoted
On 2021/7/17 上午9:45, pepperpoint@mb.ardentcoding.com wrote:
quoted
Hello,

I see this message on dmesg:

[ 2452.256756] BTRFS critical (device dm-0): corrupt leaf:

root=363 block=174113599488 slot=9 ino=7415, invalid nlink:

has 2 expect no more than 1 for dir

[ 2452.256776] BTRFS error (device dm-0): block=174113599488

read time tree block corruption detected
Please provide the following dump:

btrfs ins dump-tree -b 174113599488 /dev/dm-0
=============================================

Thanks,

Qu
quoted
When I run btrfs scrub and btrfs check, no error was detected.

I am running Linux 5.12.15-arch1-1 and btrfs-progs v5.12.1

How should I fix this error?
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help