2 unstable releases
0.2.0 | Dec 7, 2024 |
---|---|
0.1.0 | Dec 29, 2023 |
#108 in Authentication
160 downloads per month
1.5MB
26K
SLoC
webauthn_rp
webauthn_rp
is a library for server-side
Web Authentication (WebAuthn) Relying Party
(RP) operations.
The purpose of a server-side RP library is to be modular so that any client can be used with it as a backend
including native applications—WebAuthn technically only covers web applications; however it's relatively easy
to adapt to native applications as well. It achieves this by not assuming how data is sent to/from the client;
having said that, there are pre-defined serialization formats for "common" deployments which can be used when
serde
is enabled.
Cargo "features"
custom
or both bin
and serde
must be enabled; otherwise a compile_error
will occur.
bin
Enables binary (de)serialization via Encode
and Decode
. Since registered credentials will almost always
have to be saved to persistent storage, some form of (de)serialization is necessary. In the event bin
is
unsuitable or only partially suitable (e.g., human-readable output is desired), one will need to enable
custom
to allow construction of certain types (e.g., AuthenticatedCredential
).
If possible and desired, one may wish to save the data "directly" to avoid any potential temporary allocations.
For example StaticState::encode
will return a Vec
containing hundreds (and possibly thousands in the
extreme case) of bytes if the underlying public key is an RSA key. This additional allocation and copy of data
is obviously avoided if StaticState
is stored as a
composite type or its fields are stored in separate
columns when written to a relational database (RDB).
custom
Exposes functions (e.g., AuthenticatedCredential::new
) that allows one to construct instances of types that
cannot be constructed when bin
or serde
is not enabled.
serde
Enables (de)serialization of data sent to/from the client via serde
based on the JSON-motivated definitions (e.g.,
RegistrationResponseJSON
). Since
data has to be sent to/from the client, some form of (de)serialization is necessary. In the event serde
is unsuitable or only partially suitable, one will need to enable custom
to allow construction
of certain types (e.g., Registration
).
Code is strongly encouraged to rely on the Deserialize
implementations as much as possible to reduce the
chances of improperly deserializing the client data.
Note that clients are free to send data in whatever form works best, so there is no requirement the JSON-motivated definitions are used even when JSON is sent. This is especially relevant since the JSON-motivated definitions were only added in WebAuthn Level 3; thus many deployments only partially conform. Some specific deviations that may require partial customization of deserialization are the following:
ArrayBuffer
s encoded using something other than base64url.ArrayBuffer
s that are encoded multiple times (including the use of different encodings each time).- Missing fields (e.g.,
transports
). - Different field names (e.g.,
extensions
instead ofclientExtensionResults
).
serde_relaxed
Automatically enables serde
in addition to "relaxed" Deserialize
implementations
(e.g., RegistrationRelaxed
). Roughly "relaxed" translates to unknown fields being ignored and only
the fields necessary for construction of the type are required. Case still matters, duplicate fields are still
forbidden, and interrelated data validation is still performed when applicable. This can be useful when one
wants to accommodate non-conforming clients or clients that implement older versions of the spec.
serializable_server_state
Automatically enables bin
in addition to Encode
and Decode
implementations for
RegistrationServerState
and AuthenticationServerState
. Less accurate SystemTime
is used instead of
Instant
for timeout enforcement. This should be enabled if you don't desire to use in-memory collections to
store the instances of those types.
Note even when written to persistent storage, an application should still periodically remove expired ceremonies.
If one is using a relational database (RDB); then one can achieve this by storing ServerState::sent_challenge
,
the Vec
returned from Encode::encode
, and ServerState::expiration
and periodically remove all rows
whose expiration exceeds the current date and time.
Registration and authentication
Both registration and authentication ceremonies rely on "challenges", and these challenges are inherently temporary. For this reason the data associated with challenge completion can often be stored in memory without concern for out-of-memory (OOM) conditions. There are several benefits to storing such data in memory:
- No data manipulation
- By leveraging move semantics, the data sent to the client cannot be mutated once the ceremony begins.
- Improved timeout enforcement
- By ensuring the same machine that started the ceremony is also used to finish the ceremony, deviation of
system clocks is not a concern. Additionally, allowing serialization requires the use of some form of
cross-platform "timestamp" (e.g., Unix time) which differ in
implementation (e.g., platforms implement leap seconds in different ways) and are often not monotonically
increasing. If data resides in memory, a monotonic
Instant
can be used instead.
- By ensuring the same machine that started the ceremony is also used to finish the ceremony, deviation of
system clocks is not a concern. Additionally, allowing serialization requires the use of some form of
cross-platform "timestamp" (e.g., Unix time) which differ in
implementation (e.g., platforms implement leap seconds in different ways) and are often not monotonically
increasing. If data resides in memory, a monotonic
It is for those reasons data like RegistrationServerState
are not serializable by default and require the
use of in-memory collections (e.g., FixedCapHashSet
). To better ensure OOM is not a concern, RPs should set
reasonable timeouts. Since ceremonies can only be completed by moving data (e.g.,
RegistrationServerState::verify
), ceremony completion is guaranteed to free up the memory used—
RegistrationServerState
instances are only 48 bytes on x86_64-unknown-linux-gnu
platforms. To avoid issues
related to incomplete ceremonies, RPs can periodically iterate the collection for expired ceremonies and remove
such data. Other techniques can be employed as well to mitigate OOM, but they are application specific and
out-of-scope. If this is undesirable, one can enable serializable_server_state
so that RegistrationServerState
and AuthenticationServerState
implement Encode
and Decode
. Another
reason one may need to store this information persistently is for load-balancing purposes where the server that
started the ceremony is not guaranteed to be the server that finishes the ceremony.
Supported signature algorithms
The only supported signature algorithms are the following:
- Ed25519 as defined in RFC 8032 § 5.1. This corresponds
to
CoseAlgorithmIdentifier::Eddsa
. - ECDSA as defined in SEC 1 Version 2.0 § 4.1 using SHA-256
as the hash function and NIST P-256 as defined in
NIST SP 800-186 § 3.2.1.3
for the underlying elliptic curve. This corresponds to
CoseAlgorithmIdentifier::Es256
. - ECDSA as defined in SEC 1 Version 2.0 § 4.1 using SHA-384 as the hash function and NIST P-384 as defined in
NIST SP 800-186 § 3.2.1.4
for the underlying elliptic curve. This corresponds to
CoseAlgorithmIdentifier::Es384
. - RSASSA-PKCS1-v1_5 as defined in RFC 8017 § 8.2 using
SHA-256 as the hash function. This corresponds to
CoseAlgorithmIdentifier::Rs256
.
Correctness of code
This library more strictly adheres to the spec than many other similar libraries including but not limited to the following ways:
- CTAP2 canonical CBOR encoding form.
Deserialize
implementations requiring exact conformance (e.g., not allowing unknown data).- More thorough interrelated data validation (e.g., all places a Credential ID exists must match).
- Implement a lot of recommended (i.e., SHOULD) criteria (e.g., User display names conforming to the Nickname Profile as defined in RFC 8266).
Unfortunately like almost all software, this library has not been formally verified; however great care is employed in the following ways:
- Leverage move semantics to prevent mutation of data once in a static state.
- Ensure a great many invariants via types.
- Reduce code duplication.
- Reduce variable mutation allowing for simpler algebraic reasoning.
panic
-free code[^note] (i.e., define true/total functions).- Ensure arithmetic "side effects" don't occur (e.g., overflow).
- Aggressive use of compiler and Clippy lints.
- Unit tests for common cases, edge cases, and error cases.
Cryptographic libraries
This library does not rely on any sensitive data (e.g., private keys) as only signature verification is
ever performed. This means that the only thing that matters with the libraries used is their algorithmic
correctness and not other normally essential aspects like susceptibility to side-channel attacks. While I
personally believe the libraries that are used are at least as "secure" as alternatives even when dealing with
sensitive data, one only needs to audit the correctness of the libraries to be confident in their use. In fact
curve25519_dalek
has been formally
verified when the fiat
backend is used making it objectively
better than many other libraries whose correctness has not been proven. Two additional benefits of the library
choices are simpler APIs making it more likely their use is correct and better cross-platform compatibility.
Minimum Supported Rust Version (MSRV)
This will frequently be updated to be the same as stable. Specifically, any time stable is updated and that update has "useful" features or compilation no longer succeeds (e.g., due to new compiler lints), then MSRV will be updated.
MSRV changes will correspond to a SemVer patch version bump pre-1.0.0
; otherwise a minor version bump.
SemVer Policy
- All on-by-default features of this library are covered by SemVer
- MSRV is considered exempt from SemVer as noted above
License
Licensed under either of
- Apache License, Version 2.0 (LICENSE-APACHE)
- MIT license (LICENSE-MIT)
at your option.
Contribution
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.
Before any PR is sent, cargo clippy
and cargo t
should be run for each possible combination of "features"
using stable Rust. One easy way to achieve this is by building ci
and invoking it with no commands in the
webauthn_rp
directory or sub-directories. You can fetch ci
via git clone https://git.philomathiclife.com/repos/ci
,
and it can be built with cargo build --release
. Additionally,
RUSTDOCFLAGS="--cfg docsrs" cargo +nightly doc --all-features
should be run to ensure documentation can be built.
Status
This package is actively maintained and will conform to the latest WebAuthn API version. Previous versions will not be supported—excluding bug fixes of course—however functionality will exist to facilitate the migration process from the previous version.
The crate is only tested on x86_64-unknown-linux-gnu
and x86_64-unknown-openbsd
targets, but it should work
on most platforms.
[^note]: panic
s related to memory allocations or stack overflow are possible since such issues are not
formally guarded against.
Dependencies
~14MB
~265K SLoC