aboutsummaryrefslogtreecommitdiff
path: root/README
blob: fd68598c0f71fce47533ea280ad0a5c2ad35e133 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90

 About
 *****

This directory contains the source code of a collection of tools for working
with SquashFS file systems.

The `gensquashfs` program takes as input a file listing similar to the
program `gen_init_cpio` in the Linux kernel source tree and produces a
SquashFS image.

The input list precisely specifies the directory structure, all permission bits
and all UIDs & GIDs of who owns what.

The tool doesn't care if the directories, symlinks, device special files,
etc... actually exist when packing a SquashFS image. The only thing that
really has to exist are the input _files_ that can be placed arbitrarily in
the file system by specifying input and target locations.

An SELinux labeling file can be specified to add SELinux tags.

All directory entries are sorted by name and processed sequentially. All time
stamps in the SquashFS image are set to a command line specified value (or 0
by default). Thus the entire process should be deterministic, i.e. same input
produces byte-for-byte the same output.


The `rdsquashfs` program can read a SquashFS image and produce file listings,
extract individual files or entire sub trees of the image to a desired
location.


 Why not use the official squashfs-tools?
 ****************************************

The mksquashfs utility is semi-broken and generally a PITA to work with.

For the typical use case of SquashFS (i.e. as rootfs for a live distro or an
embedded system), it should be blindingly obvious that I might want to micro
manage what goes into the file system, that UIDs/GIDs of the host system are
garbage inside the image and that setting the desired permissions (e.g. suid)
or SELinux labels on the input is completely out of the question. Also, it
would be really cool if the whole thing was reproducible.

All of this seems to have been an afterthought with mksquashfs. Some of it can
be achieved with exclusion options, "pseudo files" or "filters" (a completely
undocumented feature that not even `--help` tells you about).

My main gripes with mksquashfs were the following:

 - I need to precisely replicate the entire filesystem for packing, even tough
   the only thing actually needed are in theory the regular files.
 - Files in the input FS but not in the pseudo file are still packed
   but with garbage UID/GID from the host system.
 - When I want files that are not owned by root, the root inode will get
   garbage UID/GID from the host system and there is no way to change this.
 - mtime is read from the input file system and there is no way to override it.
 - Data is packed by a thread pool, i.e. in a non-deterministic way.
 - Extended attributes are read from the input file system, i.e. the only way
   to get SELinux labels into the SquashFS filesystem is to set them on the
   input data.

That's at least what I can think of right now from the top of my head.

It would be preferable to fix mksquashfs itself, but the source code is a
horrid dumpster fire. It turned out to be easier to understand the structure
of SquashFS by reading the available documentation plus kernel source and
implementing the desired feature set from scratch.

Furthermore, upstream seems to be unmaintained at the moment and the mailing
list appears to be about as dead as SourceForge that hosts it.


 Limitations
 ***********

At the moment, the following things still require some work:

 - documentation
 - testing
 - extended attributes
    - currently limited to SELinux labeling only
    - rdsquashfs ignores them entirely
    - SquashFS supports deduplicating values through "out of line"
      storage but this is currently not used yet.
    - empty directories cannot have xattrs. The way I understand it, this is a
      design flaw in SquashFS. I hope I'm missing something here.
 - sparse files (not implemented yet)
 - hard links (not implemented yet; do we even want this?)
 - NFS export tables (not implemented yet)