security.capability xattr missing after btrfs send/receive
I found that the security.capability xattr gets lost when cloning a subvolume via btrfs send | btrfs receive. This is easy to reproduce (currently on 3.19.0-rc6, x86_64, with btrfs-progs 3.18.1):
Have two btrfs file systems, "/fs1" and "/fs2".
# btrfs sub create /fs1/test # touch /fs1/test/foo # setcap cap_net_admin+ep /fs1/test/foo # getcap /fs1/test/foo /fs1/test/foo = cap_net_admin+ep # btrfs prop set /fs1/test ro true # btrfs send /fs1/test | btrfs receive /fs2 # getcap /fs2/test/foo [reports nothing]
The cause of the problem seems to be that BTRFS_SEND_C_SET_XATTR comes before BTRFS_SEND_C_CHMOD in the send stream, so that the receiving end will first correctly restore all the xattrs of the file, and after that, perform an lchown() which will cause the kernel to reset the file's capabilities.
I don't know how to fix this correctly (especially in the case of incremental sends), so I don't have a patch, but figured it might be useful to share what I found.
Regards,
Jan
Signature · 1 line
Jan Andres <[email protected]>
Quoted 23 lines · Jan Andres
I found that the security.capability xattr gets lost when cloning a subvolume via btrfs send | btrfs receive. This is easy to reproduce (currently on 3.19.0-rc6,
x86_64, with btrfs-progs 3.18.1):Have two btrfs file systems, "/fs1" and "/fs2".
# btrfs sub create /fs1/test # touch /fs1/test/foo # setcap cap_net_admin+ep /fs1/test/foo # getcap /fs1/test/foo /fs1/test/foo = cap_net_admin+ep # btrfs prop set /fs1/test ro true # btrfs send /fs1/test | btrfs receive /fs2 # getcap /fs2/test/foo [reports nothing]The cause of the problem seems to be that
BTRFS_SEND_C_SET_XATTRcomes beforeBTRFS_SEND_C_CHMODin the send stream, so that the receiving end will first correctly restore all the xattrs of the file, and after that, perform anlchown()which will cause the kernel to reset the file's capabilities.
Quoted 3 lines
I don't know how to fix this correctly (especially in the case of incremental sends), so I don't have a patch, but figured it might be useful to share what I found.
https://bugzilla.kernel.org/show_bug.cgi?id=68891
Signature · 2 lines
Juan Orti https://miceliux.com
Looks like security.capabilities also has tripped up tar.
https://bugzilla.redhat.com/show_bug.cgi?id=771927
There's a suggestion that security.capabilities are "non standard attributes"? Also, comment 28 is interesting. Anyway it seems a big messy to support them in any case.