Package Metadata for Core Files

Intended audience: hackers working on userspace subsystems that create ELF binaries or parse ELF core files.

Motivation

ELF binaries get stamped with a unique, build-time generated hex string identifier called build-id, which gets embedded as an ELF note called .note.gnu.build-id. In most cases, this allows a stripped binary to be associated with its debugging information. It is used, for example, to dynamically fetch DWARF symbols from a debuginfo server, or to query the local package manager and find out the package metadata or, again, the DWARF symbols or program sources.

However, this usage of the build-id requires either local metadata, usually set up by the package manager, or access to a remote server over the network. Both of those might be unavailable or forbidden.

Thus it becomes desirable to add additional metadata to a binary at build time, so that systemd-coredump and other services analyzing core files are able to extract said metadata simply from the core file itself, without external dependencies.

Implementation

This document will attempt to define a common metadata format specification, so that multiple implementers might use it when building packages, or core file analyzers, and so on.

The metadata will be embedded in a single, new, 4-bytes-aligned, allocated, 0-padded, read-only ELF header section, in a name-value JSON object format. Implementers working on parsing core files should not assume a specific list of names, but parse anything that is included in the section, and should look for the note using the note type. Implementers working on build tools should strive to use the same names, for consistency. The most common will be listed here. When corresponding to the content of os-release, the values should match, again for consistency.

If available, the metadata should also include the debuginfod server URL that can provide the original executable, debuginfo and sources, to further facilitate debugging.

• Section header

SECTION: `.note.package`
note type: `0xcafe1a7e`
Owner: `FDO` (FreeDesktop.org)
Value: a single JSON object encoded as a zero-terminated UTF-8 string

• JSON payload

{
     "type":"rpm",          # this provides a namespace for the package+package-version fields
     "os":"fedora",
     "osVersion":"33",
     "name":"coreutils",
     "version":"4711.0815.fc13",
     "architecture":"arm32",
     "osCpe": "cpe:/o:fedoraproject:fedora:33",          # A CPE name for the operating system, `CPE_NAME` from os-release is a good default
     "debugInfoUrl": "https://debuginfod.fedoraproject.org/"
}

The format is a single JSON object, encoded as a zero-terminated UTF-8 string. Each name in the object shall be unique as per recommendations of RFC8259. Strings shall not contain any control character, nor use \uXXX escaping.

When it comes to JSON numbers, this specification assumes that JSON parsers processing this information are capable of reproducing the full signed 53bit integer range (i.e. -2⁵³+1…+2⁵³-1) as well as the full 64-bit IEEE floating point number range losslessly (with the exception of NaN/-inf/+inf, since JSON cannot encode that), as per recommendations of RFC8259. Fields in these JSON objects are thus permitted to encode numeric values from these ranges as JSON numbers, and should not use numeric values not covered by these types and ranges.

Reference implementations of packaging tools for .deb and .rpm are available, and provide macros/helpers to include the note in binaries built by the package build system. They make use of the new --package-metadata flag that is available in the bfd, gold, mold and lld linkers (versions 2.39, 1.3.0 and 15.0 respectively). This linker flag takes a JSON payload as parameter.

Well-known keys

The metadata format is intentionally left open, so that vendors can add their own information. A set of well-known keys is defined here, and hopefully shared among all vendors.