Skip to contentSkip to conversation

Errors on newly created file system

DISCUSSION Ferry Toth 3 participants 12 messages Apr 20, 2025 – Apr 26, 2025 Btrfs file system

lore
Share & tools
lore

Ferry Toth

author
#0 lore raw

The following is originally done by Yocto's bitbake, but when I try manually it reproduces.

I create a new fs on  a file using -r as ordinary user, then btrfs check the file (before or after mounting makes no difference), also as an ordinary user.

The fs has 1000's of errors, I cut most because it seems the same type of errors. The files system is unrepaired bootable, but can be repaired using --repair, in which 1000's of files are moved to lost+found.

The below was mkfs on a non-existing file, but writing to 16GB dduped file (rootfs is 1.4GB) makes no difference. Neither does dropping --shrink, -m or -n.

Also, writing the file to an actual disk and then check the disk gives the same errors.

What could this be?

ferry@delfion:~/tmp/edison/edison-scarthgap$ mkfs.btrfs -n 4096 --shrink -M -v -r /home/ferry/tmp/edison-intel/my/edison-morty/out/linux64/build/tmp/work/edison-poky-linux/edison-image/1.0/rootfs edison-image-edison.rootfs.btrfs
btrfs-progs v6.6.3
See https://btrfs.readthedocs.io for more information.

ERROR: zoned: unable to stat edison-image-edison.rootfs.btrfs
NOTE: several default settings have changed in version 5.15, please make sure this does not affect your deployments:
      - DUP for metadata (-m dup)
      - enabled no-holes (-O no-holes)
      - enabled free-space-tree (-R free-space-tree)

Rootdir from: /home/ferry/tmp/edison-intel/my/edison-morty/out/linux64/build/tmp/work/edison-poky-linux/edison-image/1.0/rootfs Shrink:           yes
Label:              (null)
UUID:               c2ecfaca-168a-401b-a12a-e73694d7485a
Node size:          4096
Sector size:        4096
Filesystem size:    1.43GiB
Block group profiles:
  Data+Metadata:    single            1.42GiB
  System:           single            4.00MiB
SSD detected:       no
Zoned device:       no
Incompat features:  mixed-bg, extref, skinny-metadata, no-holes, free-space-tree
Runtime features:   free-space-tree
Checksum:           crc32c
Number of devices:  1
Devices:
   ID        SIZE  PATH
    1     1.43GiB  edison-image-edison.rootfs.btrfs

ferry@delfion:~/tmp/edison/edison-scarthgap$ btrfs check edison-image-edison.rootfs.btrfs
Opening filesystem to check...
Checking filesystem on edison-image-edison.rootfs.btrfs
UUID: c2ecfaca-168a-401b-a12a-e73694d7485a
[1/7] checking root items
[2/7] checking extents
[3/7] checking free space tree
[4/7] checking fs roots
root 5 inode 252551099 errors 2000, link count wrong
        unresolved ref dir 260778488 index 2 namelen 11 name COPYING.MIT filetype 1 errors 0
root 5 inode 252551611 errors 2000, link count wrong
        unresolved ref dir 260775820 index 79 namelen 18 name generic_Apache-2.0 filetype 1 errors 0
root 5 inode 252575069 errors 2000, link count wrong
        unresolved ref dir 260777976 index 3 namelen 10 name recipeinfo filetype 1 errors 0
root 5 inode 252575538 errors 2000, link count wrong
        unresolved ref dir 260778506 index 3 namelen 10 name recipeinfo filetype 1 errors 0
root 5 inode 252577241 errors 2000, link count wrong
        unresolved ref dir 260777972 index 3 namelen 10 name recipeinfo filetype 1 errors 0
root 5 inode 256713617 errors 2000, link count wrong
        unresolved ref dir 260776096 index 3 namelen 12 name GPL-2.0-only filetype 1 errors 0
root 5 inode 256713619 errors 2000, link count wrong
        unresolved ref dir 260776096 index 5 namelen 13 name LGPL-2.1-only filetype 1 errors 0
root 5 inode 256713620 errors 2000, link count wrong
        unresolved ref dir 260776096 index 6 namelen 10 name recipeinfo filetype 1 errors 0
root 5 inode 256730804 errors 2000, link count wrong
        unresolved ref dir 260777541 index 3 namelen 7 name COPYING filetype 1 errors 0
        unresolved ref dir 260777543 index 3 namelen 7 name COPYING filetype 1 errors 0
root 5 inode 256730805 errors 2000, link count wrong
        unresolved ref dir 260777541 index 4 namelen 11 name COPYING.LIB filetype 1 errors 0
        unresolved ref dir 260777543 index 4 namelen 11 name COPYING.LIB filetype 1 errors 0
root 5 inode 256730806 errors 2000, link count wrong
        unresolved ref dir 260777541 index 5 namelen 15 name COPYING.RUNTIME filetype 1 errors 0
        unresolved ref dir 260777543 index 5 namelen 15 name COPYING.RUNTIME filetype 1 errors 0
root 5 inode 256730807 errors 2000, link count wrong
        unresolved ref dir 260777541 index 6 namelen 8 name COPYING3 filetype 1 errors 0
        unresolved ref dir 260777543 index 6 namelen 8 name COPYING3 filetype 1 errors 0
root 5 inode 256730808 errors 2000, link count wrong
        unresolved ref dir 260777541 index 7 namelen 12 name COPYING3.LIB filetype 1 errors 0
        unresolved ref dir 260777543 index 7 namelen 12 name COPYING3.LIB filetype 1 errors 0

...

        unresolved ref dir 260777434 index 4 namelen 10 name recipeinfo filetype 1 errors 0 unresolved ref dir 260777436 index 4 namelen 10 name recipeinfo filetype 1 errors 0 unresolved ref dir 260777438 index 4 namelen 10 name recipeinfo filetype 1 errors 0 root 5 inode 260775819 errors 2000, link count wrong
        unresolved ref dir 260775820 index 2 namelen 16 name license.manifest filetype 1 errors 0
ERROR: errors found in fs roots
found 1061572608 bytes used, error(s) found
total csum bytes: 990472
total tree bytes: 47329280
total fs tree bytes: 44457984
total extent tree bytes: 1511424
btree space waste bytes: 11573386
file data blocks allocated: 1014243328
 referenced 1014243328

#1 lore raw
Quoted 45 lines · Ferry Toth

The following is originally done by Yocto's bitbake, but when I try manually it reproduces.

I create a new fs on  a file using -r as ordinary user, then btrfs check the file (before or after mounting makes no difference), also as an ordinary user.

The fs has 1000's of errors, I cut most because it seems the same type of errors. The files system is unrepaired bootable, but can be repaired using --repair, in which 1000's of files are moved to lost+found.

The below was mkfs on a non-existing file, but writing to 16GB dduped file (rootfs is 1.4GB) makes no difference. Neither does dropping -- shrink, -m or -n.

Also, writing the file to an actual disk and then check the disk gives the same errors.

What could this be?

ferry@delfion:~/tmp/edison/edison-scarthgap$ mkfs.btrfs -n 4096 --shrink 
-M -v -r /home/ferry/tmp/edison-intel/my/edison-morty/out/linux64/build/ 
tmp/work/edison-poky-linux/edison-image/1.0/rootfs edison-image- edison.rootfs.btrfs
btrfs-progs v6.6.3
See https://btrfs.readthedocs.io for more information.

ERROR: zoned: unable to stat edison-image-edison.rootfs.btrfs
NOTE: several default settings have changed in version 5.15, please make sure
      this does not affect your deployments:
      - DUP for metadata (-m dup)
      - enabled no-holes (-O no-holes)
      - enabled free-space-tree (-R free-space-tree)

Rootdir from: /home/ferry/tmp/edison-intel/my/edison-morty/out/linux64/ build/tmp/work/edison-poky-linux/edison-image/1.0/rootfs
  Shrink:           yes
Label:              (null)
UUID:               c2ecfaca-168a-401b-a12a-e73694d7485a
Node size:          4096
Sector size:        4096
Filesystem size:    1.43GiB
Block group profiles:
  Data+Metadata:    single            1.42GiB
  System:           single            4.00MiB
SSD detected:       no
Zoned device:       no
Incompat features:  mixed-bg, extref, skinny-metadata, no-holes, free- space-tree
Runtime features:   free-space-tree
Checksum:           crc32c
Number of devices:  1
Devices:
   ID        SIZE  PATH
    1     1.43GiB  edison-image-edison.rootfs.btrfs

Quoted 11 lines
ferry@delfion:~/tmp/edison/edison-scarthgap$ btrfs check edison-image- edison.rootfs.btrfs
Opening filesystem to check...
Checking filesystem on edison-image-edison.rootfs.btrfs
UUID: c2ecfaca-168a-401b-a12a-e73694d7485a
[1/7] checking root items
[2/7] checking extents
[3/7] checking free space tree
[4/7] checking fs roots
root 5 inode 252551099 errors 2000, link count wrong
         unresolved ref dir 260778488 index 2 namelen 11 name 
COPYING.MIT filetype 1 errors 0

Looks like exactly the nlink bugs related to --rootdir option.

And that's fixed in v6.10 first, by the commit c6464d3f99ed ("btrfs-progs: mkfs: rework how we traverse rootdir"), then further improved in v6.12 with the commit ef1157473372 ("btrfs-progs: mkfs: add hard link support for --rootdir").

So in short, if the directory contains hardlinks out of the directory, then you have to use btrfs-progs newer than v6.12.

Thanks,
Qu

author
#2 lore raw

Hi,

Op 21-04-2025 om 00:00 schreef Qu Wenruo:

Quoted 57 lines · Ferry Toth
Quoted 43 lines

The following is originally done by Yocto's bitbake, but when I try manually it reproduces.

I create a new fs on  a file using -r as ordinary user, then btrfs check the file (before or after mounting makes no difference), also as an ordinary user.

The fs has 1000's of errors, I cut most because it seems the same type of errors. The files system is unrepaired bootable, but can be repaired using --repair, in which 1000's of files are moved to lost+found.

The below was mkfs on a non-existing file, but writing to 16GB dduped file (rootfs is 1.4GB) makes no difference. Neither does dropping -- shrink, -m or -n.

Also, writing the file to an actual disk and then check the disk gives the same errors.

What could this be?

ferry@delfion:~/tmp/edison/edison-scarthgap$ mkfs.btrfs -n 4096 -- shrink -M -v -r /home/ferry/tmp/edison-intel/my/edison-morty/out/ linux64/build/ tmp/work/edison-poky-linux/edison-image/1.0/rootfs edison-image- edison.rootfs.btrfs
btrfs-progs v6.6.3
See https://btrfs.readthedocs.io for more information.

ERROR: zoned: unable to stat edison-image-edison.rootfs.btrfs
NOTE: several default settings have changed in version 5.15, please make sure this does not affect your deployments:
       - DUP for metadata (-m dup)
       - enabled no-holes (-O no-holes)
       - enabled free-space-tree (-R free-space-tree)

Rootdir from: /home/ferry/tmp/edison-intel/my/edison-morty/out/ linux64/ build/tmp/work/edison-poky-linux/edison-image/1.0/rootfs Shrink:           yes
Label:              (null)
UUID:               c2ecfaca-168a-401b-a12a-e73694d7485a
Node size:          4096
Sector size:        4096
Filesystem size:    1.43GiB
Block group profiles:
   Data+Metadata:    single            1.42GiB
   System:           single            4.00MiB
SSD detected:       no
Zoned device:       no
Incompat features:  mixed-bg, extref, skinny-metadata, no-holes, free- space-tree
Runtime features:   free-space-tree
Checksum:           crc32c
Number of devices:  1
Devices:
    ID        SIZE  PATH
     1     1.43GiB  edison-image-edison.rootfs.btrfs

Quoted 10 lines
ferry@delfion:~/tmp/edison/edison-scarthgap$ btrfs check edison-image- edison.rootfs.btrfs
Opening filesystem to check...
Checking filesystem on edison-image-edison.rootfs.btrfs
UUID: c2ecfaca-168a-401b-a12a-e73694d7485a
[1/7] checking root items
[2/7] checking extents
[3/7] checking free space tree
[4/7] checking fs roots
root 5 inode 252551099 errors 2000, link count wrong
         unresolved ref dir 260778488 index 2 namelen 11 name COPYING.MIT filetype 1 errors 0
Quoted 6 lines

Looks like exactly the nlink bugs related to --rootdir option.

And that's fixed in v6.10 first, by the commit c6464d3f99ed ("btrfs- progs: mkfs: rework how we traverse rootdir"), then further improved in v6.12 with the commit ef1157473372 ("btrfs-progs: mkfs: add hard link support for --rootdir").

So in short, if the directory contains hardlinks out of the directory, then you have to use btrfs-progs newer than v6.12.

Hi Qu I am confirming v6.12 resolves this issue. Afaik Yocto uses reflinks. I'm guessing that generates the same issue?

As a work around for people on Ubuntu Noble (20.04 LTS) with btrfs-progs version 6.6.3-1.1build2, installing the package from Plucky (no other dependencies) with version 6.12-1build1 solves this issue.

Yocto users on Scarthgap (5.0 LTS) with version 6.7.1 may copy the recipe meta/recipes-devtools/btrfs-tools/btrfs-tools_6.13.bb from walnascar or 6.14 from master. If they are building additional tools that use headers from this package like btrfs-compsize these may break.

Quoted 2 lines

Thanks,
Qu

#3 lore raw
Quoted 60 lines · Ferry Toth

Hi,

Op 21-04-2025 om 00:00 schreef Qu Wenruo:

Quoted 57 lines · Ferry Toth
Quoted 53 lines

The following is originally done by Yocto's bitbake, but when I try manually it reproduces.

I create a new fs on  a file using -r as ordinary user, then btrfs check the file (before or after mounting makes no difference), also as an ordinary user.

The fs has 1000's of errors, I cut most because it seems the same type of errors. The files system is unrepaired bootable, but can be repaired using --repair, in which 1000's of files are moved to lost+found.

The below was mkfs on a non-existing file, but writing to 16GB dduped file (rootfs is 1.4GB) makes no difference. Neither does dropping -- shrink, -m or -n.

Also, writing the file to an actual disk and then check the disk gives the same errors.

What could this be?

ferry@delfion:~/tmp/edison/edison-scarthgap$ mkfs.btrfs -n 4096 -- shrink -M -v -r /home/ferry/tmp/edison-intel/my/edison-morty/out/ linux64/build/ tmp/work/edison-poky-linux/edison-image/1.0/rootfs edison-image- edison.rootfs.btrfs
btrfs-progs v6.6.3
See https://btrfs.readthedocs.io for more information.

ERROR: zoned: unable to stat edison-image-edison.rootfs.btrfs
NOTE: several default settings have changed in version 5.15, please make sure this does not affect your deployments:
       - DUP for metadata (-m dup)
       - enabled no-holes (-O no-holes)
       - enabled free-space-tree (-R free-space-tree)

Rootdir from: /home/ferry/tmp/edison-intel/my/edison-morty/out/ linux64/ build/tmp/work/edison-poky-linux/edison-image/1.0/rootfs Shrink:           yes
Label:              (null)
UUID:               c2ecfaca-168a-401b-a12a-e73694d7485a
Node size:          4096
Sector size:        4096
Filesystem size:    1.43GiB
Block group profiles:
   Data+Metadata:    single            1.42GiB
   System:           single            4.00MiB
SSD detected:       no
Zoned device:       no
Incompat features:  mixed-bg, extref, skinny-metadata, no-holes, free- space-tree
Runtime features:   free-space-tree
Checksum:           crc32c
Number of devices:  1
Devices:
    ID        SIZE  PATH
     1     1.43GiB  edison-image-edison.rootfs.btrfs

ferry@delfion:~/tmp/edison/edison-scarthgap$ btrfs check edison- image- edison.rootfs.btrfs
Opening filesystem to check...
Checking filesystem on edison-image-edison.rootfs.btrfs
UUID: c2ecfaca-168a-401b-a12a-e73694d7485a
[1/7] checking root items
[2/7] checking extents
[3/7] checking free space tree
[4/7] checking fs roots
root 5 inode 252551099 errors 2000, link count wrong
         unresolved ref dir 260778488 index 2 namelen 11 name COPYING.MIT filetype 1 errors 0
Quoted 8 lines
Quoted 6 lines

Looks like exactly the nlink bugs related to --rootdir option.

And that's fixed in v6.10 first, by the commit c6464d3f99ed ("btrfs- progs: mkfs: rework how we traverse rootdir"), then further improved in v6.12 with the commit ef1157473372 ("btrfs-progs: mkfs: add hard link support for --rootdir").

So in short, if the directory contains hardlinks out of the directory, then you have to use btrfs-progs newer than v6.12.

Hi Qu I am confirming v6.12 resolves this issue. Afaik Yocto uses reflinks. I'm guessing that generates the same issue?

Reflink should not cause the problem, shared extents (reflinks) are not longer shared inside the new btrfs.

E.g. if two inodes shares the same data extent, it will be created as two data extents, one for each inode. Thus it will cause extra space usage.

It's only the hard links causing problems, as older progs directly uses the st_nlink reported from the host fs, but it's very possible that, some hard links are not inside the rootdir, thus causing missing nlinks inside the created btrfs.

Quoted 2 lines

As a work around for people on Ubuntu Noble (20.04 LTS) with btrfs-progs version 6.6.3-1.1build2, installing the package from Plucky (no other dependencies) with version 6.12-1build1 solves this issue.

Sorry we didn't notice the bug early enough, thus the fixes are only landed into v6.12, and we do not maintain backports for older progs.

Thus I'm afraid Yocto or similar projects have to require a much newer progs instead.

Thanks,
Qu

Quoted 6 lines

Yocto users on Scarthgap (5.0 LTS) with version 6.7.1 may copy the recipe meta/recipes-devtools/btrfs-tools/btrfs-tools_6.13.bb from walnascar or 6.14 from master. If they are building additional tools that use headers from this package like btrfs-compsize these may break.

Quoted 2 lines

Thanks,
Qu

author
#4 lore raw

Op woensdag 23 april 2025 00:00:36 CEST schreef Qu Wenruo:

Quoted 107 lines · Ferry Toth
Quoted 85 lines

Hi,

Op 21-04-2025 om 00:00 schreef Qu Wenruo:

Quoted 82 lines · Ferry Toth
Quoted 58 lines

The following is originally done by Yocto's bitbake, but when I try manually it reproduces.

I create a new fs on a file using -r as ordinary user, then btrfs check the file (before or after mounting makes no difference), also as an ordinary user.

The fs has 1000's of errors, I cut most because it seems the same type of errors. The files system is unrepaired bootable, but can be repaired using --repair, in which 1000's of files are moved to lost+found.

The below was mkfs on a non-existing file, but writing to 16GB dduped file (rootfs is 1.4GB) makes no difference. Neither does dropping -- shrink, -m or -n.

Also, writing the file to an actual disk and then check the disk gives the same errors.

What could this be?

ferry@delfion:~/tmp/edison/edison-scarthgap$ mkfs.btrfs -n 4096 -- 
shrink -M -v -r /home/ferry/tmp/edison-intel/my/edison-morty/out/ 
linux64/build/ tmp/work/edison-poky-linux/edison-image/1.0/rootfs 
edison-image- edison.rootfs.btrfs
btrfs-progs v6.6.3
See https://btrfs.readthedocs.io for more information.

ERROR: zoned: unable to stat edison-image-edison.rootfs.btrfs
NOTE: several default settings have changed in version 5.15, please make sure
this does not affect your deployments:
- DUP for metadata (-m dup)
- enabled no-holes (-O no-holes)
- enabled free-space-tree (-R free-space-tree)

Rootdir from: /home/ferry/tmp/edison-intel/my/edison-morty/out/ linux64/ build/tmp/work/edison-poky-linux/edison-image/1.0/rootfs
Shrink: yes

Label:              (null)
UUID:               c2ecfaca-168a-401b-a12a-e73694d7485a
Node size:          4096
Sector size:        4096
Filesystem size:    1.43GiB
Block group profiles:
   Data+Metadata:    single            1.42GiB
   System:           single            4.00MiB
SSD detected:       no
Zoned device:       no
Incompat features:  mixed-bg, extref, skinny-metadata, no-holes, 
free- space-tree
Runtime features:   free-space-tree
Checksum:           crc32c
Number of devices:  1

Devices:
ID SIZE PATH
1 1.43GiB edison-image-edison.rootfs.btrfs

Quoted 12 lines
ferry@delfion:~/tmp/edison/edison-scarthgap$ btrfs check edison- 
image- edison.rootfs.btrfs
Opening filesystem to check...
Checking filesystem on edison-image-edison.rootfs.btrfs
UUID: c2ecfaca-168a-401b-a12a-e73694d7485a
[1/7] checking root items
[2/7] checking extents
[3/7] checking free space tree
[4/7] checking fs roots
root 5 inode 252551099 errors 2000, link count wrong
         unresolved ref dir 260778488 index 2 namelen 11 name 
COPYING.MIT filetype 1 errors 0

Looks like exactly the nlink bugs related to --rootdir option.

And that's fixed in v6.10 first, by the commit c6464d3f99ed ("btrfs-
progs: mkfs: rework how we traverse rootdir"), then further improved in v6.12 with the commit ef1157473372 ("btrfs-progs: mkfs: add hard link support for --rootdir").

Quoted 5 lines
Quoted 2 lines

So in short, if the directory contains hardlinks out of the directory, then you have to use btrfs-progs newer than v6.12.

Hi Qu I am confirming v6.12 resolves this issue. Afaik Yocto uses reflinks. I'm guessing that generates the same issue?

Reflink should not cause the problem, shared extents (reflinks) are not longer shared inside the new btrfs.

E.g. if two inodes shares the same data extent, it will be created as two data extents, one for each inode.
Thus it will cause extra space usage.

It's only the hard links causing problems, as older progs directly uses the st_nlink reported from the host fs, but it's very possible that, some hard links are not inside the rootdir, thus causing missing nlinks inside the created btrfs.

Quoted 1 line
Quoted 6 lines
Quoted 3 lines

As a work around for people on Ubuntu Noble (20.04 LTS) with btrfs-progs version 6.6.3-1.1build2, installing the package from Plucky (no other dependencies) with version 6.12-1build1 solves this issue.

Sorry we didn't notice the bug early enough, thus the fixes are only landed into v6.12, and we do not maintain backports for older progs.

I never noticed either, while I was in a very good position to do so.

Times when I noticed filesystem errors I attributed these to unstable kernel crash. Which was probably true, never noticed errors at T0 until now.

Quoted 2 lines

Thus I'm afraid Yocto or similar projects have to require a much newer progs instead.

Yes that is the workaround below. You can copy the newer ”recipe” and that overrides the older one. Of course goes all the way and brings in new headers which can break dependencies (btrfs-compsize in this case).

Quoted 10 lines

Thanks,
Qu

Quoted 7 lines

Yocto users on Scarthgap (5.0 LTS) with version 6.7.1 may copy the recipe meta/recipes-devtools/btrfs-tools/btrfs-tools_6.13.bb from walnascar or 6.14 from master. If they are building additional tools that use headers from this package like btrfs-compsize these may break.

Quoted 2 lines

Thanks,
Qu

While here, am I right that we can not generate the rootfs with compression on?

Reason I ask is, Yocto of course builds the rootfs and than has mkfs.btrfs create the image. But it runs as unprivileged user, so can not do mount. And then can not do defrag.

Quoted 4 lines
Quoted 2 lines
#5 lore raw
Quoted 2 lines · Ferry Toth

Op woensdag 23 april 2025 00:00:36 CEST schreef Qu Wenruo:

[...]

Quoted 14 lines
Quoted 5 lines
Quoted 4 lines

Yocto users on Scarthgap (5.0 LTS) with version 6.7.1 may copy the

recipe meta/recipes-devtools/btrfs-tools/btrfs-tools_6.13.bb from

Quoted 7 lines
Quoted 7 lines

walnascar or 6.14 from master. If they are building additional tools

that use headers from this package like btrfs-compsize these may break.

Quoted 3 lines

Thanks,

Qu

Quoted 6 lines

While here, am I right that we can not generate the rootfs with compression on?

Reason I ask is, Yocto of course builds the rootfs and than has mkfs.btrfs create the image. But it runs as unprivileged user, so can not do mount.

And then can not do defrag.

We have this feature recently thanks to Mark!

In the latest release v6.13, there is a new option "--compress" added to mkfs.btrfs, which must be used with "--rootdir".

And the result is exactly what you expected, mkfs.btrfs will try to compress the file extents at runtime. For uncompressible data, it will detect at the beginning and fallback to uncompressed data instead, exactly like the kernel.

But considering how new this feature this, it will be appreciated if Yocto guys can do some extra testing to make sure nothing is broken. (Normally a btrfs check after the mkfs will be good enough)

Thanks,
Qu

Quoted 12 lines
Quoted 9 lines
Quoted 5 lines
author
#6 lore raw

Op donderdag 24 april 2025 01:24:23 CEST schreef Qu Wenruo:

Quoted 32 lines · Ferry Toth
Quoted 2 lines

Op woensdag 23 april 2025 00:00:36 CEST schreef Qu Wenruo:

[...]

Quoted 18 lines
Quoted 5 lines
Quoted 4 lines

Yocto users on Scarthgap (5.0 LTS) with version 6.7.1 may copy the

recipe meta/recipes-devtools/btrfs-tools/btrfs-tools_6.13.bb from

Quoted 7 lines
Quoted 7 lines

walnascar or 6.14 from master. If they are building additional tools

that use headers from this package like btrfs-compsize these may break.

Quoted 3 lines

Thanks,

Qu

While here, am I right that we can not generate the rootfs with compression on?

Quoted 5 lines

Reason I ask is, Yocto of course builds the rootfs and than has mkfs.btrfs create the image. But it runs as unprivileged user, so can not do mount.

And then can not do defrag.

We have this feature recently thanks to Mark!

Quoted 7 lines

In the latest release v6.13, there is a new option "--compress" added to mkfs.btrfs, which must be used with "--rootdir".

And the result is exactly what you expected, mkfs.btrfs will try to compress the file extents at runtime.
For uncompressible data, it will detect at the beginning and fallback to uncompressed data instead, exactly like the kernel.

That is great, thanks!

Quoted 4 lines

But considering how new this feature this, it will be appreciated if Yocto guys can do some extra testing to make sure nothing is broken. (Normally a btrfs check after the mkfs will be good enough)

I will test this.

In the meanwhile I found another problem (and I am not the first it seems see /https://stackoverflow.com/questions/79475262/yocto-root-filesystem-ownership-issue/79590289#79590289/[1])

Yocto uses pseudo to fake root ownership. Even though the directory to be copied into the new fs is owned by an unprivileged user inside the btrfs image the files are owned by root, when created by mkfs.ext4 and mkfs.btrfs.

Except with newer mkfs.btrfs (I tested using 6.13) the files are owned by the unprivileged user.

The result is, the image will not boot correctly.

Quoted 18 lines

Thanks,
Qu

Quoted 12 lines
Quoted 9 lines
Quoted 5 lines

--------
[1] https://stackoverflow.com/questions/79475262/yocto-root-filesystem-ownership-issue/79590289#79590289

author
#7 lore raw

Hi

Op 24-04-2025 om 13:31 schreef Ferry Toth:

Quoted 107 lines

Op donderdag 24 april 2025 01:24:23 CEST schreef Qu Wenruo:

Quoted 8 lines · Ferry Toth
Quoted 4 lines

Op woensdag 23 april 2025 00:00:36 CEST schreef Qu Wenruo:

Quoted 7 lines

[...]

Quoted 3 lines

  > > Yocto users on Scarthgap (5.0 LTS) with version 6.7.1 may

copy the

Quoted 7 lines
Quoted 7 lines

  > > recipe meta/recipes-devtools/btrfs-tools/btrfs-tools_6.13.bb from

  > > walnascar or 6.14 from master. If they are building

additional tools

Quoted 3 lines
Quoted 3 lines

  > > that use headers from this package like btrfs-compsize these

may break.

Quoted 32 lines
Quoted 22 lines

  > >> Thanks,

  > >> Qu

While here, am I right that we can not generate the rootfs with

compression on?

Reason I ask is, Yocto of course builds the rootfs and than has

Quoted 7 lines

mkfs.btrfs create the image. But it runs as unprivileged user, so can

not do mount.

And then can not do defrag.

Quoted 5 lines

We have this feature recently thanks to Mark!

In the latest release v6.13, there is a new option "--compress"

added to

Quoted 9 lines

mkfs.btrfs, which must be used with "--rootdir".

And the result is exactly what you expected, mkfs.btrfs will try to

compress the file extents at runtime.

For uncompressible data, it will detect at the beginning and

fallback to

Quoted 1 line

uncompressed data instead, exactly like the kernel.

That is great, thanks!

Quoted 7 lines

But considering how new this feature this, it will be appreciated if

Yocto guys can do some extra testing to make sure nothing is broken.

(Normally a btrfs check after the mkfs will be good enough)

I will test this.

In the meanwhile I found another problem (and I am not the first it seems see

/https://stackoverflow.com/questions/79475262/yocto-root-filesystem-ownership-issue/79590289#79590289/ <https://stackoverflow.com/questions/79475262/yocto-root-filesystem-ownership-issue/79590289#79590289>)

Quoted 9 lines

Yocto uses pseudo to fake root ownership. Even though the directory to be copied into the new fs is owned by an unprivileged user inside the btrfs image the files are owned by root, when created by mkfs.ext4 and mkfs.btrfs.

Except with newer mkfs.btrfs (I tested using 6.13) the files are owned by the unprivileged user.

The result is, the image will not boot correctly.

I found more about this issue here

https://lore.kernel.org/yocto/[email protected]/T/#m5de0afa17d2c0f640e86ffe67e0d74aea467fd5b

Quoted 38 lines
Quoted 35 lines

Thanks,

Qu

Quoted 23 lines

  > >

  > >

  >

  >

#8 lore raw
Quoted 2 lines · Ferry Toth

Hi

[...]

Quoted 5 lines
Quoted 5 lines

Except with newer mkfs.btrfs (I tested using 6.13) the files are owned by the unprivileged user.

Quoted 7 lines
Quoted 3 lines

The result is, the image will not boot correctly.

I found more about this issue here

https://lore.kernel.org/yocto/ [email protected]/T/ #m5de0afa17d2c0f640e86ffe67e0d74aea467fd5b

Thanks for the report.

Just want to be sure, with pseudo emulating root environment, how does it handle the file uid/gid?

Mkfs.btrfs uses the uid/gid reported from stat() system calls, thus if pseudo doesn't change uid/gid reported from stat(), mkfs.btrfs will just follow the result.

I guess it's possible for us to implement an idmap-like solution, but I'd like to know how pseudo works first.

Thanks,
Qu

author
#9 lore raw

Op 24-04-2025 om 22:49 schreef Qu Wenruo:

Quoted 16 lines · Ferry Toth
Quoted 2 lines

Hi

[...]

Quoted 10 lines
Quoted 8 lines

Except with newer mkfs.btrfs (I tested using 6.13) the files are owned by the unprivileged user.

The result is, the image will not boot correctly.

I found more about this issue here

Quoted 7 lines
Quoted 2 lines

https://lore.kernel.org/yocto/ [email protected]/T/ #m5de0afa17d2c0f640e86ffe67e0d74aea467fd5b

Thanks for the report.

Just want to be sure, with pseudo emulating root environment, how does it handle the file uid/gid?

Mkfs.btrfs uses the uid/gid reported from stat() system calls, thus if pseudo doesn't change uid/gid reported from stat(), mkfs.btrfs will just follow the result.

The change in mkfs.btrfs is reportedly:
"The important change is that previously mkfs.btrfs was using lstat syscall to get file stats, but then it switched to nftw libc function to do the same. Pseudo[3] however doesn't support this call - it "knows" that it exists (at least nftw64), but there seems to be no implementation."

So, as I understand, pseudo would hook into stats call, lookup in database and change uid/gid accordingly. But now there is only a stub.

Right now I am trying a patch for pseudo that implements nftw64. It triggers a rebuild for half my image, so that will takes another hour or so. I'll let you know how that goes.

This particular problem does seem to be more pseudo related.

Quoted 4 lines

I guess it's possible for us to implement an idmap-like solution, but I'd like to know how pseudo works first.

Thanks,
Qu

author
#10 lore raw

Hi

Op 24-04-2025 om 23:16 schreef Ferry Toth:

Quoted 32 lines

Op 24-04-2025 om 22:49 schreef Qu Wenruo:

Quoted 16 lines · Ferry Toth
Quoted 2 lines

Hi

[...]

Quoted 10 lines
Quoted 8 lines

Except with newer mkfs.btrfs (I tested using 6.13) the files are owned by the unprivileged user.

The result is, the image will not boot correctly.

I found more about this issue here

Quoted 7 lines
Quoted 2 lines

https://lore.kernel.org/yocto/ [email protected]/T/ #m5de0afa17d2c0f640e86ffe67e0d74aea467fd5b

Thanks for the report.

Just want to be sure, with pseudo emulating root environment, how does it handle the file uid/gid?

Mkfs.btrfs uses the uid/gid reported from stat() system calls, thus if pseudo doesn't change uid/gid reported from stat(), mkfs.btrfs will just follow the result.

The change in mkfs.btrfs is reportedly:
"The important change is that previously mkfs.btrfs was using lstat syscall to get file stats, but then it switched to nftw libc function to do the same. Pseudo[3] however doesn't support this call - it "knows" that it exists (at least nftw64), but there seems to be no implementation."

Quoted 5 lines

So, as I understand, pseudo would hook into stats call, lookup in database and change uid/gid accordingly. But now there is only a stub.

Right now I am trying a patch for pseudo that implements nftw64. It triggers a rebuild for half my image, so that will takes another hour or so. I'll let you know how that goes.

This particular problem does seem to be more pseudo related.

Yes, the patch for pseudo is working. So, this is not a mkfs.btrfs bug.

Thanks for your support!

Quoted 5 lines
Quoted 4 lines

I guess it's possible for us to implement an idmap-like solution, but I'd like to know how pseudo works first.

Thanks,
Qu

author
#11 lore raw
In reply to Qu Wenruo

Op 24-04-2025 om 01:24 schreef Qu Wenruo:

Quoted 6 lines · Ferry Toth
Quoted 5 lines

Op woensdag 23 april 2025 00:00:36 CEST schreef Qu Wenruo:

While here, am I right that we can not generate the rootfs with compression on?

Quoted 7 lines
Quoted 3 lines

Reason I ask is, Yocto of course builds the rootfs and than has mkfs.btrfs create the image. But it runs as unprivileged user, so can not do mount.

And then can not do defrag.

We have this feature recently thanks to Mark!

In the latest release v6.13, there is a new option "--compress" added to mkfs.btrfs, which must be used with "--rootdir".

I built 6.14 while configuring with --enable-lzo

Configure says:
zstd support: no
lzo support: yes

Finally running mkfs.btrfs with --compress lzo

I get:
-EXPERIMENTAL -INJECT -STATIC -LZO -ZSTD -UDEV +FSVERITY -ZONED CRYPTO=builtin

And
ERROR: lzo support not compiled in

Strangely: when I run with --compress zlib and I check the resulting image with btrfs-compsize it says:

Processed 8969984 files, 2574779 regular extents (6008702 refs), 5007593 inline.
Type Perc Disk Usage Uncompressed Referenced

TOTAL       91%      532G         582G         820G
none       100%      489G         489G         575G
zlib        38%      350M         902M         926M
lzo         46%       42G          92G         243G
prealloc   100%       99M          99M         678M

Seems like it tries lzo anyway and is mixed up about support not built in.