This system has been merged into OpenBSD base. If you'd like to contribute to rpki-client, please mail your patches to tech@openbsd.org. This repository is simply the OpenBSD version plus some glue for portability. It is updated from time to time to keep in sync with OpenBSD's version.
rpki-client is an implementation of RPKI (resource public key infrastructure) described in RFC 6480. It implements the client side of RPKI, which is responsible for downloading and validating route origin statements. For usage, please read rpki-client(8).
The design focus of rpki-client is simplicity and security. To wit, it implements RPKI components necessary for validating route statements and omits superfluities (such as, for example, which X509 certificate sections must be labelled "Critical").
The system runs on most modern UNIX operating systems with the the
OpenSSL external library installed, version
1.1.1c and above.
Beyond that it requires only
BSD make, usually called bmake
on systems
where GNU make is the default.
See Portability for details.
The reference operating system is OpenBSD, which we strongly suggest for all installations for security reasons. It will support LibreSSL once the library gains CMS parsing.
It has been tested on OpenBSD, FreeBSD, Linux (glibc and musl), and IllumOS. Solaris, NetBSD, and Mac OS X require some portability work.
See the TODO file for open questions regarding RPKI operation in general.
rpki-client is written as part of the rpki-client(8) project, an RPKI validator for OpenBSD. It was funded by NetNod, IIS.SE, SUNET and 6connect.
First, you'll need a recent OpenSSL library
(version 1.1.1c and above) on your operating system. At this point, just
run the following. The installation rule will install into PREFIX
,
defaulting to /usr/local, which you may override:
% ./configure PREFIX=/opt/local
% make
# make install
If your pkg-config
for OpenSSL 1.1.1c and above isn't openssl
(or
eopenssl11
for OpenBSD), pass the proper name as a build option.
For example, using openssl111
:
% ./configure
% make PKG_OPENSSL=openssl111
This value may also be hard-coded in the Makefile. If your system consistently uses a different package name, please raise an issue to let us know.
Next, you'll need the /var/cache/rpki-client directory in place. It must be writable by the operator of rpki-client. The default output directory is /var/db/rpki-client, which must also be writable (if not overriden).
You'll also need TAL ("trust anchor locator") files. There are some in the tals directory of this system, but you can download them on your own. For default operation, load these into /etc/rpki.
These default paths are set as
RPKI_PATH_BASE_DIR
, RPKI_PATH_OUT_DIR
, and RPKI_TAL_DIR
,
respectively, in the
Makefile.
Alternatively, override the variables when invoking make
, e.g.,
% make RPKI_TAL_DIR=/etc/tals
You'll also need openrsync(1) or
rsync as specified with the -e argument.
To hardcode an alternate rsync implementation, set the
RPKI_RSYNC_COMMAND
value in the
Makefile.
In the following, the first uses a custom TAL file, while the second loads all TAL files from their default location. Output for the first is written into ./openbgpd and /var/db/rpki-client/openbgpd for the second.
% ./rpki-client -v -t sometal.tal .
% ./rpki-client -v
If you later want to uninstall the system, simply run
# make uninstall
If the manpages in the install root have already been indexed, you may need to re-run makewhatis(8) to purge the system's manpage.
The rpki-client run-time is split into at least three processes which pass data back and forth. "At least" since the system will dynamically spawn additional process in addition to the three core processes. Most of the architecture is implemented in main.c.
The master process orchestrates all other process. It also formats and outputs valid route data.
The first subordinate process is responsible for obtaining certificates, route announcements, manifests, and so on. It waits for the master process to give it a repository and destination, then executes openrsync(1) and waits for termination. It executes child openrsync(1) processes asynchronously for maximum efficiency.
Side note: although rpki-client can use rsync instead of openrsync(1), this is not recommended for security reasons: the latter has been designed to make maximum use of OpenBSD's security framework, as has rpki-client.
The second subordinate process parses and validates data files. It is given filenames by the master process, parses them in-order, and returns the results. The files are assumed to exist on disc by virtue of being downloaded earlier by the first subordinate process. This process performs the bulk of the work.
The master process is responsible for orchestrating this pipeline. It seeds the parser process with the TAL files, retrieves TAL output, then begins parsing certificates (X509), manifests (MFT), revocation lists (CRL), and Route Origin Authorizations (ROAs). If any of these files sits in a repository not yet fetched, that repository is fetched or refreshed. When the repository is fetched, those pending entries are flushed into the parser.
The master process also outputs valid routes. At this time, it does so in bgpd.conf(5), BIRD, RIPE NCC RPKI JSON, or CSV formats.
It's trivially possible to put each type of file parse into its own process, but it's not clear whether this adds any security since the file-system available to a parser consists of all file types.
Alternatively, each repository might have its own parser that's restricted to files only within the repository. This would allow unveil(2) to limit the parser only to those in its repository. The repository cache would need to be redesigned to nest repositories so that a top-level repository would be able to access its children.
The latter is not difficult to implement.
At its heart, rpki-client is a tool for validating hierarchical statements. The terminals of this hierarchy consist of IP address prefix and numerical AS identifier components. The non-terminal statements provide both acceptable ranges of both components and links to further terminal and non-terminal nodes in the tree.
Terminal nodes are ROA (route origin authorisation) and CRL (certificate revocation list) files. Non-terminal nodes are X509 (certificate) and MFT (manifest) files. The root node (there may be multiple roots) is a TAL (trust anchor locator) file.
The validation algorithm is a breadth-first (though whether depth or breadth first is irrelevant) tree walk.
Most of the certificate validation in RPKI comes from the X509_STORE
functionality of OpenSSL. This covers CRL revocation, expiration dates,
and so on.
It begins by parsing a TAL file, RFC 7730, which specifies a trust anchor certificate address and its public key. The parsing and validation of the TAL file occurs in tal.c.
Side note: the TAL file may technically specify multiple top-level certificates; but in the case of rpki-client, only the first is processed.
A trust anchor is an X509 (RFC 6487) certificate given by the TAL file. Beyond the usual certificate parsing in cert.c, the trust anchor files also have a number of additional constraints imposed in validate.c:
- the certificate must be self-signed
- the public key must match the one given in the TAL file
- it must have an SKI (subject key identifier)
- the SKI must be unique in the set of all parsed certificates (trust anchors and otherwise)
- must not specify a CRL resource
Furthermore:
- it may only contain non-inheritance AS identifiers
- it may only contain non-inheritance IP blocks
Each trust anchor (inheriting from the X509 validation) contains a reference to a manifest file that's used for further parsing.
Manifests (RFC 6486) contain links to more resources. They are parsed in mft.c, with the CMS (RFC 6488) envelope parsed in cms.c, and additional checks implemented in validate.c.
- self-signed CMS envelope
- CMS envelope self-signed certificate is signed by the AKI's certificate
- manifest time window has not expired
Manifests contain a list of files they manage that must be ROA, CRL, or
X509 (roa
, crl
, or cer
suffixes, respectively).
Each file is associated with a hash.
Stale manifests---those whose validity period has elapsed---are accepted (and noted), but will contain zero members.
ROA (route origin authorisation, RFC 6482) files are stipulated in manifests. These are the focus of RPKI: those that pass validation are emitted as valid routes. ROA files consist of data wrapped in a CMS envelope. They are parsed in roa.c, with the CMS (RFC 6488) envelope parsed in cms.c, and additional checks implemented in validate.c.
- computed digest matches that given by the manifest
- self-signed CMS envelope
- CMS envelope self-signed certificate is signed by the AKI's certificate
- IP blocks must be within the ranges allocated by the nearest non-inheriting certificate in the chain to the trust anchor
An ROA may technically contain zero IP prefixes. If this is the case, it is merely skipped.
A "stale" ROA (time validity has elapsed) is also ignored.
X509 certificates (RFC 6487) certificate are the mainstay of RPKI's validation. They are parsed in cert.c with further validation being performed in validate.c.
- computed digest matches that given by the manifest (if applicable)
- the certificate must be signed by the AKI's certificate
- the SKI must be unique in the set of all parsed certificates (trust anchors and otherwise)
- must specify a CRL resource
- AS identifiers/ranges must be within the ranges allocated by the nearest non-inheriting certificate in the chain to the trust anchor (see TODO for notes)
- IP blocks must be within the ranges allocated by the nearest non-inheriting certificate in the chain to the trust anchor
rpki-client imposes no specific checks on CRL than those provided by
OpenSSL's X509_STORE
functionality.
Some repositories, however, contain enormous CRL files with thousands and thousands of entries. These take quite some time to parse.
The rpki-client is portable to the extent that it will compile and run on most modern UNIX systems. To date it is known to compile on GNU/Linux (musl, glibc), FreeBSD, OpenBSD, and IllumOS (OmniOS). Portability efforst are underway to NetBSD, Darwin (Mac OS X), and Solaris, all of which are missing ppoll(2) and in the latter case, mkostemp.
It uses oconfigure for its compatibility layer and minci for continuous integration.
However, the system depends heavily on OpenBSD's security mechanisms (only enabled on OpenBSD installations) to safely and securely parse untrusted content. Those running on a non-OpenBSD operating system should be aware that this additional protection is not available.
If the RPKI_PRIVDROP
macro evaluates to 1 as set in the
Makefile,
the RPKI_PRIVDROP_USER
is used as the username into which to
privilege-drop.
On OpenBSD, this is _rpki-client.
Privilege dropping only occurs when running the utility as root.
If RPKI_PRIVDROP
is set to 0, no privilege dropping occurs.
rpki-client makes significant use of pledge(2) to constrain resources available to the running process. On FreeBSD, the same (or similar) may be effected by judicious use of Capsicum. On Linux, seccomp, although it's an unholy mess.
This function is used in main.c. On non-OpenBSD systems it is redefined to be empty in extern.h.
Once TAL files have been parsed (these may sit anywhere on the file-system), the parsing process restricts file-system access to the local repository directory with unveil(2).
It's not trivial to port this to FreeBSD or Linux.
First, calls to BIO_new_file
would need to use BIO_new_fp
with a
separate fdopen
call.
This descriptor would need to be opened with openat
and the input
paths stripped of their common prefix.
This way, calls directly to open
could be filtered.
This function is used in main.c. On non-OpenBSD systems it is redefined to be empty in extern.h.