Re: [gentoo-dev] [pre-GLEP r3] Gentoo binary package container format

[Michał Górny]

Here's the newest version.

Changes:

- added explicit notion of parent directory (missing in previous GLEP
but present in implementation),

- explicitly named GNU tar format with list of permitted extensions,

- changed volume label to 'gpkg-1.txt' file to improve portability; made
it explicit version identifier as well,

- added info on other package formats to rationale.


---
GLEP: 9999
Title: Gentoo binary package container format
Author: Michał Górny <mgorny@g.o>
Type: Standards Track
Status: Draft
Version: 1
Created: 2018-11-15
Last-Modified: 2018-11-26
Post-History: 2018-11-17
Content-Type: text/x-rst
---

Abstract
========

This GLEP proposes a new binary package container format for Gentoo.
The current tbz2/XPAK format is shortly described, and its deficiences
are explained.  Accordingly, the requirements for a new format are set
and a gpkg format satisfying them is proposed.  The rationale for
the design decisions is provided.


Motivation
==========

The current Portage binary package format
-----------------------------------------

The historical ``.tbz2`` binary package format used by Portage is
a concatenation of two distinct formats: header-oriented compressed .tar
format (used to hold package files) and trailer-oriented custom XPAK
format (used to hold metadata)  [#MAN-XPAK]_.  The format has already
been extended incompatibly twice.

The first time, support for storing multiple successive builds of binary
package for a single ebuild version has been added.  This feature relies
on appending additional hyphen, followed by an integer to the package
filename.  It is disabled by default (preserving backwards
compatibility) and controlled by ``binpkg-multi-instance`` feature.

The second time, support for additional compression formats has been
added.  When format other than bzip2 is used, the ``.tbz2`` suffix
is replaced by ``.xpak`` and Portage relies on magic bytes to detect
compression used.  For backwards compatibility, Portage still defaults
to using bzip2; compression program can be switched using
``BINPKG_COMPRESS`` configuration variable.

Additionally, there have been minor changes to the stored metadata
and file storage policies.  In particular, behavior regarding
``INSTALL_MASK``, controllable file compression and stripping has
changed over time.


The advantages of tbz2/XPAK format
----------------------------------

The tbz2/XPAK format used by Portage has three interesting features:

1. **Each binary package is fully contained within a single file.**
   While this might seem unnecessary, it makes it easier for the user
   to transfer binary packages without having to be concerned about
   finding all the necessary files to transfer.

2. **The binary packages are compatible with regular compressed
   tarballs, most of the time.**  With notable exceptions of historical
   versions of pbzip2 and the recent zstd compressor, tbz2/XPAK packages
   can be extracted using regular tar utility with a compressor
   implementation that discards trailing garbage.

3. **The metadata is uncompressed, and can be efficiently accessed
   without decompressing package contents.**  This includes
   the possibility of rewriting it (e.g. as a result of package moves)
   without the necessity of repacking the files.


Transparency problem with the current binary package format
-----------------------------------------------------------

Notwithstanding its advantages, the tbz2/XPAK format has a significant
design fault that consists of two issues:

1. **The XPAK format is a custom binary format with explicit use
   of binary-encoded file offsets and field lengths.**  As such, it is
   non-trivial to read or edit without specialized tools.  Such tools
   are currently implemented separately from the package manager,
   as part of the portage-utils toolkit, written in C [#PORTAGE-UTILS]_.

2. **The tarball compatibility feature relies on obscure feature of
   ignoring trailing garbage in compressed files**.  While this is
   implemented consistently in most of the compressors, this feature
   is not really a part of specification but rather traditional
   behavior.  Given that the original reasons for this no longer apply,
   new compressor implementations are likely to miss support for this.

Both of the issues make the format hard to use without dedicated tools,
or when the tools misbehave.  This impacts the following scenarios:

A. **Using binary packages for system recovery.**  In case of serious
   breakage, it is really preferable that the format depends on as few
   tools a possible, and especially not on Gentoo-specific tools.

B. **Inspecting binary packages in detail exceeding standard package
   manager facilities.**

C. **Modifying binary packages in ways not predicted by the package
   manager authors.**  A real-life example of this is working around
   broken ``pkg_*`` phases which prevent the package from being
   installed.


OpenPGP extensibility problem
-----------------------------

There are at least three obvious ways in which the current format could
be extended to support OpenPGP signatures, and each of them has its own
distinct problem:

1. **Adding a detached signature.**  This option is non-intrusive but
   causes the format to no longer be contained in a single file.

2. **Wrapping the package in OpenPGP message format.**  This would use
   a standard format and make verification and unpacking relatively
   easy.  However, it would break backwards compatibility and add
   explicit dependency on OpenPGP implementation in order to unpack
   the package.

3. **Adding OpenPGP signature as extra XPAK member.**  This is
   the clever solution.  It implies strengthening the dependency
   on custom tooling, now additionally necessary to extract
   the signature and reconstruct the original file to accommodate
   verification.


Goals for a new container format
--------------------------------

All of the above considered, the new format should combine
the advantages of the existing format and at the same time address its
deficiencies whenever possible.  Furthermore, since a format replacement
is taking place it is worthwhile to consider additional goals that could
be satisfied with little change.

The following obligatory goals have been set for a replacement format:

1. **The packages must remain contained in a single file.**  As a matter
   of user convenience, it should be possible to transfer binary
   packages without having to use multiple files, and to install them
   from any location.

2. **The file format must be entirely based on common file formats,
   respecting best practices, with as little customization as necessary
   to satisfy the requirements.**  The format should be transparent
   enough to let user inspect and manipulate it without special tooling
   or detailed knowledge.

3. **The file format must provide support for OpenPGP signatures.**
   Preferably, it should use standard OpenPGP message formats.

4. **The file format must allow for efficient metadata updates.**
   In particular, it should be possible to update the metadata without
   having to recompress package files.

Additionally, the following optional goals have been noted:

A. **The file format should account for easy recognition both through
   filename and through contents.**  Preferably, it should have distinct
   features making it possible to detect it via file(1).

B. **The file format should provide for partial fetching of binary
   packages.**  It should be possible to easily fetch and read
   the package metadata without having to download the whole package.

C. **The file format should allow for metadata compression.**

D. **The file format should make future extensions easily possible
   without breaking backwards compatibility.**


Specification
=============

The container format
--------------------

The gpkg package container is an uncompressed .tar achive whose filename
should use ``.gpkg.tar`` suffix.  This archive contains the following
members, all placed in a single directory whose name matches
the basename of the package file, in order:

1. The package identifier file ``gpkg-1.txt`` (required).

2. A signature for the metadata archive: ``metadata.tar${comp}.sig``
   (optional).

3. The metadata archive ``metadata.tar${comp}``, optionally compressed
   (required).

4. A signature for the filesystem image archive:
   ``image.tar${comp}.sig`` (optional).

5. The filesystem image archive ``image.tar${comp}``, optionally
   compressed (required).

It is recommended that relative order of the archive members is
preserved.  However, implementations must support archives with members
out of order.

The container may be extended with additional members in the future.
The implementations should ignore unrecognized members and preserve
them across package updates.


Permitted .tar format features
------------------------------

The tar archives should use either the POSIX ustar format or a subset
of the GNU format with the following (optional) extensions:

- long pathnames and long linknames,

- base-256 encoding of large file sizes.

Other extensions should be avoided whenever possible.


The package identifier file
---------------------------

The package identifier file serves the purpose of identifying the binary
package format and its version.

The implementations must include a package identifier file named
``gpkg-1.txt``.  The filename includes package format version;
implementations should reject packages which do not contain this file
as unsupported format.

The file can have any contents.  Normally, it should be empty.

Furthermore, this file should be included in the .tar archive
as the first member.  This makes it possible to use it as an additional
magic at a fixed location that can be used by tools such as file(1)
to easily distinguish Gentoo binary packages from regular .tar archives.


The metadata archive
--------------------

The metadata archive stores the package metadata needed for the package
manager to process it.  The archive should be included at the beginning
of the binary package in order to make it possible to read it out of
partially fetched binary package, and to avoid fetching the remaining
part of the package if not necessary.

The archive contains a single directory called ``metadata``.  In this
directory, the individual metadata keys are stored as files.  The exact
keys and metadata format is outside the scope of this specification.

The package manager may need to modify the package metadata.  In this
case, it should replace the metadata archive without having to alter
other package members.

The metadata archive can optionally be compressed.  It can also be
supplemented with a detached OpenPGP signature.


The image archive
-----------------

The image archive stores all the files to be installed by the binary
package.  It should be included as the last of the files in the binary
package container.

The archive contains a single directory called ``image``.  Inside this
directory, all package files are stored in filesystem layout, relative
to the root directory.

The image archive can optionally be compressed.  It can also be
supplemented with a detached OpenPGP signature.


Archive member compression
--------------------------

The archive members outlined above support optional compression using
one of the compressed file formats supported by the package manager.
The exact list of compression types is outside the scope of this
specification.

The implementations must support archive members being uncompressed,
and must support using different compression types for different files.

When compressing an archive member, the member filename should be
suffixed using the standard suffix for the particular compressed file
type (e.g. ``.bz2`` for bzip2 format).


OpenPGP member signatures
-------------------------

The archive members support optional OpenPGP signatures.
The implementations must allow the user to specify whether OpenPGP
signatures are to be expected in remotely fetched packages.

If the signatures are expected and the archive member is unsigned, the
package manager must reject processing it.  If the signature does not
verify, the package manager must reject processing the corresponding
archive member.  In particular, it must not attempt decompressing
compressed members in those circumstances.

The signatures are created as binary detached OpenPGP signature files,
with filename corresponding to the member filename with ``.sig`` suffix
appended.

The exact details regarding creating and verifying signatures, as well
as maintaining and distributing keys are outside the scope of this
specification.


Rationale
=========

Package formats used by other distributions
-------------------------------------------

The research on the new package format included investigating
the possibility of reusing solutions from other operating system
distributions.  While reusing a foreign package format would be
interesting, the differences in Gentoo metadata structure would prevent
any real compatibility.  Some degree of compatibility might be achieved
through adapting the Gentoo metadata, however the costs of such
a solution would probably outweigh its usefulness.

Debian and its derivates are using the .deb package format.  This is
a nested archive format, with the outer archive being of ar format,
and containing nested tarballs of control information (metadata)
and data  [#DEB-FORMAT]_.

Red Hat, its derivates and some less related distributions are using
the RPM format.  It is a custom binary format, storing metadata directly
and using a trailer cpio archive to store package files.

Arch Linux is using xz-compressed tarballs (suffixed ``.pkg.tar.xz``)
as its binary package format.  The tarballs contain package files
on top-level, with specially named dotfiles used for package metadata.
OpenPGP signatures are stored as detached ``.sig`` files alongside
packages.

Exherbo is using the pbins format.  In this format, the binary package
metadata is stored in repository alike ebuilds, and the binary package
files are stored separately and downloaded alike source tarballs.


Nested archive format
---------------------

The basic problem in designing the new format was how to embed multiple
data streams (metadata, image) into a single file.  Traditionally, this
has been done via using two non-conflicting file formats.  However,
while such a solution is clever, it suffers in terms of transparency.

Therefore, it has been established that the new format should really
consist of a single archive format, with all necessary data
transparently accessible inside the file.  Consequently, it has been
debated how different parts of binary package data should be stored
inside that archive.

The proposal to continue storing image data as top-level data
in the package format, and store metadata as special directory in that
structure has been discarded as a case of in-band signalling.

Finally, the proposal has been shaped to store different kinds of data
as nested archives in the outer binary package container.  Besides
providing a clean way of accessing different kinds of information, it
makes it possible to add separate OpenPGP signatures to them.


Inner vs. outer compression
---------------------------

One of the points in the new format debate was whether the binary
package as a whole should be compressed vs. compressing individual
members.  The first option may seem as an obvious choice, especially
given that with a larger data set, the compression may proceed more
effectively.  However, it has a single strong disadvantage: compression
prevents random access and manipulation of the binary package members.

While for the purpose of reading binary packages, the problem could be
circumvented through convenient member ordering and avoiding disjoint
reads of the binary package, metadata updates would either require
recompressing the whole package (which could be really time consuming
with large packages) or applying complex techniques such as splitting
the compressed archive into multiple compressed streams.

This considered, the simplest solution is to apply compression to
the individual package members, while leaving the container format
uncompressed.  It provides fast random access to the individual members,
as well as capability of updating them without the necessity of
recompressing other files in the container.

This also makes it possible to easily protect compressed files using
standard OpenPGP detached signature format.  All this combined,
the package manager may perform partial fetch of binary package, verify
the signature of its metadata member and process it without having to
fetch the potentially-large image part.


Container and archive formats
-----------------------------

During the debate, the actual archive formats to use were considered.
The .tar format seemed an obvious choice for the image archive since
it is the only widely deployed archive format that stores all kinds
of file metadata on POSIX systems.  However, multiple options for
the outer format has been debated.

Firstly, the ZIP format has been proposed as the only commonly supported
format supporting adding files from stdin (i.e. making it possible to
pipe the inner archives straight into the container without using
temporary files).  However, this format has been clearly rejected
as both not being present in the system set, and being trailer-based
and therefore unusable without having to fetch the whole file.

Secondly, the ar and cpio formats were considered.  The former is used
by Debian and its derivative binary packages; the latter is used by Red
Hat derivatives.  Both formats have the advantage of having less
historical baggage than .tar, and having less overhead.  However, both
are also rather obscure (especially given that ar is actually provided
by GNU binutils rather than as a stand-alone archiver), considered
obsolete by POSIX and both have file size limitations smaller than .tar.

Thirdly, SquashFS was another interesting option.  Its main advantage is
transparent compression support and ability to mount as a filesystem.
However, it has a significant implementation complexity, including mount
management and necessity of fallback to unsquashfs.  Since the image
needs to be writable for the pre-installation manipulations, using it
via a mount would additionally require some kind of overlay filesystem.
Using it as top-level format has no real gain over a pipeline with tar,
and is certainly less portable.  Therefore, there does not seem to be
a benefit in using SquashFS.

All that considered, it has been decided that there is no purpose
in using a second archive format in the specification unless it has
significant advantage to .tar.  Therefore, .tar has also been used
as outer package format, even though it has larger overhead than other
formats (mostly due to padding).


.tar portability issues
-----------------------

The modern .tar dialects could be considered a dirty extensions
of the original .tar format.  Three variants may be considered
of interest: POSIX ustar, pax (newer POSIX standard) and GNU tar.
All three formats are supported by GNU tar, whose presence on systems
used to create binary packages could be relied on.  Therefore,
the portability concerns are related mostly to being able to read
and modify binary packages in scenarios of GNU tar being unavailable.

For the purpose of this specification, a detailed research
on portability of individual tar features has been conducted.
The research concluded to:

    Judging by the test results, the most portability could be
    achieved by:

    - using strict POSIX ustar format whenever possible,

    - using GNU format for long paths (that do not fix in ustar format),

    - using base-256 (+ pax if already used) encoding for large files,

    - using pax (+ octal or base-256) for high-range/precision
      timestamps and user/group identifiers,

    - using pax attributes for extended metadata and/or volume label.

It has been determined that for the purpose of binary package we really
only need to be concerned about long paths and huge files.  Therefore,
the above was limited to the three first points and a guideline was
formed from them.

Debian has a similar guideline for the inner tar of their package
format has been created [#DEB-FORMAT]_.


Member ordering
---------------

The member ordering is explicitly specified in order to provide for
trivially reading metadata from partially fetched archives.
By requiring the metadata archive to be stored before the image archive,
the package manager may stop fetching after reading it and save
bandwidth and/or space.


Detached OpenPGP signatures
---------------------------

The use of detached OpenPGP signatures is to provide authenticity checks
for binary packages.  Covering the complete members with signatures
provide for trivial verification of all metadata and image contents
respectively, without having to invent custom mechanisms for combining
them.  Covering the compressed archives helps to prevent zipbomb
attacks.  Covering the individual members rather than the whole package
provides for verification of partially fetched binary packages.


Format versioning
-----------------

The format is versioned through an explicit file, with the version
stored in the filename.  If the format changes incompatible,
the filename changes and old implementations do not recognize it
as a valid package.

Previously, the format tried to avoid an explicit file for this purpose
and used volume label instead.  However, the use of label has been
renounced due to unforeseen portability issues.


Backwards Compatibility
=======================

The format does not preserve backwards compatibility with the tbz2
packages.  It has been established that preserving compatibility with
the old format was impossible without making the new format even worse
than the old one was.

For example, adding any visible members to the tarball would cause
them to be installed to the filesystem by old Portage versions.  Working
around this would require some kind of awful hacks that would oppose
the goal of using simple and transparent package format.


Reference Implementation
========================

The proof-of-concept implementation of binary package format converter
is available as xpak2gpkg [#XPAK2GPKG]_.  It can be used to easily
create packages in the new format for early inspection.


References
==========

.. [#MAN-XPAK] xpak - The XPAK Data Format used with Portage binary
   packages
   (https://dev.gentoo.org/~zmedico/portage/doc/man/xpak.5.html)

.. [#PORTAGE-UTILS] portage-utils: Small and fast Portage helper tools
   written in C
   (https://packages.gentoo.org/packages/app-portage/portage-utils)

.. [#DEB-FORMAT] deb(5) — Debian binary package format
   (https://manpages.debian.org/unstable/dpkg-dev/deb.5.en.html)

.. [#TAR-PORTABILITY] Michał Górny, Portability of tar features
   (https://dev.gentoo.org/~mgorny/articles/portability-of-tar-features.html)

.. [#XPAK2GPKG] xpak2gpkg: Proof-of-concept converter from tbz2/xpak
   to gpkg binpkg format
   (https://github.com/mgorny/xpak2gpkg)


Copyright
=========
This work is licensed under the Creative Commons Attribution-ShareAlike 3.0
Unported License. To view a copy of this license, visit
http://creativecommons.org/licenses/by-sa/3.0/.


-- 
Best regards,
Michał Górny