Impact

Unix-like operating systems may segfault due to dereferencing a dangling pointer in specific circumstances. This requires an environment variable to be set in a different thread than the affected functions. This may occur without the user's knowledge, notably in a third-party library.

Workarounds

No workarounds are known.

References

• time-rs/time#293

hyper's HTTP header parser accepted, according to RFC 7230, illegal contents inside Content-Length headers. Due to this, upstream HTTP proxies that ignore the header may still forward them along if it chooses to ignore the error.

To be vulnerable, hyper must be used as an HTTP/1 server and using an HTTP proxy upstream that ignores the header's contents but still forwards it. Due to all the factors that must line up, an attack exploiting this vulnerability is unlikely.

When decoding chunk sizes that are too large, hyper's code would encounter an integer overflow. Depending on the situation, this could lead to data loss from an incorrect total size, or in rarer cases, a request smuggling attack.

To be vulnerable, you must be using hyper for any HTTP/1 purpose, including as a client or server, and consumers must send requests or responses that specify a chunk size greater than 18 exabytes. For a possible request smuggling attack to be possible, any upstream proxies must accept a chunk size greater than 64 bits.

On Windows, configuring a named pipe server with pipe_mode will force ServerOptions::reject_remote_clients as false.

This drops any intended explicit configuration for the reject_remote_clients that may have been set as true previously.

The default setting of reject_remote_clients is normally true meaning the default is also overridden as false.

Workarounds

Ensure that pipe_mode is set first after initializing a ServerOptions. For example:

let mut opts = ServerOptions::new();
opts.pipe_mode(PipeMode::Message);
opts.reject_remote_clients(true);

In affected versions, after a Report is constructed using wrap_err or wrap_err_with to attach a message of type D onto an error of type E, then using downcast to recover ownership of either the value of type D or the value of type E, one of two things can go wrong:

• If downcasting to E, there remains a value of type D to be dropped. It is incorrectly "dropped" by running E's drop behavior, rather than D's. For example if D is &str and E is std::io::Error, there would be a call of std::io::Error::drop in which the reference received by the Drop impl does not refer to a valid value of type std::io::Error, but instead to &str.

• If downcasting to D, there remains a value of type E to be dropped. When D and E do not happen to be the same size, E's drop behavior is incorrectly executed in the wrong location. The reference received by the Drop impl may point left or right of the real E value that is meant to be getting dropped.

In both cases, when the Report contains an error E that has nontrivial drop behavior, the most likely outcome is memory corruption.

When the Report contains an error E that has trivial drop behavior (for example a Utf8Error) but where D has nontrivial drop behavior (such as String), the most likely outcome is that downcasting to E would leak D.

The following presentation at this year's DEF CON was brought to our attention on the SQLx Discord:

SQL Injection isn't Dead: Smuggling Queries at the Protocol Level
http://web.archive.org/web/20240812130923/https://media.defcon.org/DEF%20CON%2032/DEF%20CON%2032%20presentations/DEF%20CON%2032%20-%20Paul%20Gerste%20-%20SQL%20Injection%20Isn't%20Dead%20Smuggling%20Queries%20at%20the%20Protocol%20Level.pdf
(Archive link for posterity.)

Essentially, encoding a value larger than 4GiB can cause the length prefix in the protocol to overflow, causing the server to interpret the rest of the string as binary protocol commands or other data.

It appears SQLx does perform truncating casts in a way that could be problematic, for example: https://github.com/launchbadge/sqlx/blob/6f2905695b9606b5f51b40ce10af63ac9e696bb8/sqlx-postgres/src/arguments.rs#L163

This code has existed essentially since the beginning, so it is reasonable to assume that all published versions <= 0.8.0 are affected.

Mitigation

As always, you should make sure your application is validating untrustworthy user input. Reject any input over 4 GiB, or any input that could encode to a string longer than 4 GiB. Dynamically built queries are also potentially problematic if it pushes the message size over this 4 GiB bound.

Encode::size_hint() can be used for sanity checks, but do not assume that the size returned is accurate. For example, the Json<T> and Text<T> adapters have no reasonable way to predict or estimate the final encoded size, so they just return size_of::<T>() instead.

For web application backends, consider adding some middleware that limits the size of request bodies by default.

Resolution

sqlx 0.8.1 has been released with the fix: https://github.com/launchbadge/sqlx/blob/main/CHANGELOG.md#081---2024-08-23

Postgres users are advised to upgrade ASAP as a possible exploit has been demonstrated: https://github.com/launchbadge/sqlx/issues/3440#issuecomment-2307956901

MySQL and SQLite do not appear to be exploitable, but upgrading is recommended nonetheless.

Uint::overflowing_shl/overflowing_shr returned false-negative overflow flags. overflowing_shl missed bits shifted above BITS but within the top limb (non-limb-aligned widths such as U160), and limbs wholly discarded by shifts >= 64; overflowing_shr missed wholly discarded low limbs. Shifted values were correct; only the flag was wrong.

The wrong flag propagates: checked_shl/checked_shr return Some instead of None, strict_* fail to panic, and saturating_* return a wrapped value instead of saturating. The incorrect checked_shl result causes to_base_be (and string formatting) to loop forever on no-alloc builds for non-limb-aligned widths — a denial of service if formatting is reachable from untrusted input.

Separately, wrapping_shl/wrapping_shr on 64/128/256-bit types truncated the shift amount modulo 2^32, so shifts >= 2^32 returned an incorrectly wrapped value instead of zero; on 32-bit targets the generic path also truncated 64-bit shift amounts.

Callers using checked or saturating shift semantics on untrusted shift amounts may compute incorrect results.